Showing papers in "European Radiology in 2019"

The evolution of image reconstruction for CT—from filtered back projection to artificial intelligence

Abstract: The first CT scanners in the early 1970s already used iterative reconstruction algorithms; however, lack of computational power prevented their clinical use. In fact, it took until 2009 for the first iterative reconstruction algorithms to come commercially available and replace conventional filtered back projection. Since then, this technique has caused a true hype in the field of radiology. Within a few years, all major CT vendors introduced iterative reconstruction algorithms for clinical routine, which evolved rapidly into increasingly advanced reconstruction algorithms. The complexity of algorithms ranges from hybrid-, model-based to fully iterative algorithms. As a result, the number of scientific publications on this topic has skyrocketed over the last decade. But what exactly has this technology brought us so far? And what can we expect from future hardware as well as software developments, such as photon-counting CT and artificial intelligence? This paper will try answer those questions by taking a concise look at the overall evolution of CT image reconstruction and its clinical implementations. Subsequently, we will give a prospect towards future developments in this domain. KEY POINTS: • Advanced CT reconstruction methods are indispensable in the current clinical setting. • IR is essential for photon-counting CT, phase-contrast CT, and dark-field CT. • Artificial intelligence will potentially further increase the performance of reconstruction methods.

Medical students' attitude towards artificial intelligence: a multicentre survey

Abstract: To assess undergraduate medical students’ attitudes towards artificial intelligence (AI) in radiology and medicine. A web-based questionnaire was designed using SurveyMonkey, and was sent out to students at three major medical schools. It consisted of various sections aiming to evaluate the students’ prior knowledge of AI in radiology and beyond, as well as their attitude towards AI in radiology specifically and in medicine in general. Respondents’ anonymity was ensured. A total of 263 students (166 female, 94 male, median age 23 years) responded to the questionnaire. Around 52% were aware of the ongoing discussion about AI in radiology and 68% stated that they were unaware of the technologies involved. Respondents agreed that AI could potentially detect pathologies in radiological examinations (83%) but felt that AI would not be able to establish a definite diagnosis (56%). The majority agreed that AI will revolutionise and improve radiology (77% and 86%), while disagreeing with statements that human radiologists will be replaced (83%). Over two-thirds agreed on the need for AI to be included in medical training (71%). In sub-group analyses male and tech-savvy respondents were more confident on the benefits of AI and less fearful of these technologies. Contrary to anecdotes published in the media, undergraduate medical students do not worry that AI will replace human radiologists, and are aware of the potential applications and implications of AI on radiology and medicine. Radiology should take the lead in educating students about these emerging technologies. • Medical students are aware of the potential applications and implications of AI in radiology and medicine in general. • Medical students do not worry that the human radiologist or physician will be replaced. • Artificial intelligence should be included in medical training.

Deep learning reconstruction improves image quality of abdominal ultra-high-resolution CT

Abstract: Deep learning reconstruction (DLR) is a new reconstruction method; it introduces deep convolutional neural networks into the reconstruction flow. This study was conducted in order to examine the clinical applicability of abdominal ultra-high-resolution CT (U-HRCT) exams reconstructed with a new DLR in comparison to hybrid and model-based iterative reconstruction (hybrid-IR, MBIR). Our retrospective study included 46 patients seen between December 2017 and April 2018. A radiologist recorded the standard deviation of attenuation in the paraspinal muscle as the image noise and calculated the contrast-to-noise ratio (CNR) for the aorta, portal vein, and liver. The overall image quality was assessed by two other radiologists and graded on a 5-point confidence scale ranging from 1 (unacceptable) to 5 (excellent). The difference between CT images subjected to hybrid-IR, MBIR, and DLR was compared. The image noise was significantly lower and the CNR was significantly higher on DLR than hybrid-IR and MBIR images (p < 0.01). DLR images received the highest and MBIR images the lowest scores for overall image quality. DLR improved the quality of abdominal U-HRCT images. • The potential degradation due to increased noise may prevent implementation of ultra-high-resolution CT in the abdomen. • Image noise and overall image quality for hepatic ultra-high-resolution CT images improved with deep learning reconstruction as compared to hybrid- and model-based iterative reconstruction.

Deep learning for liver tumor diagnosis part I: development of a convolutional neural network classifier for multi-phasic MRI.

Abstract: To develop and validate a proof-of-concept convolutional neural network (CNN)–based deep learning system (DLS) that classifies common hepatic lesions on multi-phasic MRI. A custom CNN was engineered by iteratively optimizing the network architecture and training cases, finally consisting of three convolutional layers with associated rectified linear units, two maximum pooling layers, and two fully connected layers. Four hundred ninety-four hepatic lesions with typical imaging features from six categories were utilized, divided into training (n = 434) and test (n = 60) sets. Established augmentation techniques were used to generate 43,400 training samples. An Adam optimizer was used for training. Monte Carlo cross-validation was performed. After model engineering was finalized, classification accuracy for the final CNN was compared with two board-certified radiologists on an identical unseen test set. The DLS demonstrated a 92% accuracy, a 92% sensitivity (Sn), and a 98% specificity (Sp). Test set performance in a single run of random unseen cases showed an average 90% Sn and 98% Sp. The average Sn/Sp on these same cases for radiologists was 82.5%/96.5%. Results showed a 90% Sn for classifying hepatocellular carcinoma (HCC) compared to 60%/70% for radiologists. For HCC classification, the true positive and false positive rates were 93.5% and 1.6%, respectively, with a receiver operating characteristic area under the curve of 0.992. Computation time per lesion was 5.6 ms. This preliminary deep learning study demonstrated feasibility for classifying lesions with typical imaging features from six common hepatic lesion types, motivating future studies with larger multi-institutional datasets and more complex imaging appearances. • Deep learning demonstrates high performance in the classification of liver lesions on volumetric multi-phasic MRI, showing potential as an eventual decision-support tool for radiologists. • Demonstrating a classification runtime of a few milliseconds per lesion, a deep learning system could be incorporated into the clinical workflow in a time-efficient manner.

Application of a deep learning algorithm for detection and visualization of hip fractures on plain pelvic radiographs

Abstract: To identify the feasibility of using a deep convolutional neural network (DCNN) for the detection and localization of hip fractures on plain frontal pelvic radiographs (PXRs). Hip fracture is a leading worldwide health problem for the elderly. A missed diagnosis of hip fracture on radiography leads to a dismal prognosis. The application of a DCNN to PXRs can potentially improve the accuracy and efficiency of hip fracture diagnosis. A DCNN was pretrained using 25,505 limb radiographs between January 2012 and December 2017. It was retrained using 3605 PXRs between August 2008 and December 2016. The accuracy, sensitivity, false-negative rate, and area under the receiver operating characteristic curve (AUC) were evaluated on 100 independent PXRs acquired during 2017. The authors also used the visualization algorithm gradient-weighted class activation mapping (Grad-CAM) to confirm the validity of the model. The algorithm achieved an accuracy of 91%, a sensitivity of 98%, a false-negative rate of 2%, and an AUC of 0.98 for identifying hip fractures. The visualization algorithm showed an accuracy of 95.9% for lesion identification. A DCNN not only detected hip fractures on PXRs with a low false-negative rate but also had high accuracy for localizing fracture lesions. The DCNN might be an efficient and economical model to help clinicians make a diagnosis without interrupting the current clinical pathway. • Automated detection of hip fractures on frontal pelvic radiographs may facilitate emergent screening and evaluation efforts for primary physicians. • Good visualization of the fracture site by Grad-CAM enables the rapid integration of this tool into the current medical system. • The feasibility and efficiency of utilizing a deep neural network have been confirmed for the screening of hip fractures.

Deep learning-based image restoration algorithm for coronary CT angiography.

Abstract: The purpose of this study was to compare the image quality of coronary computed tomography angiography (CTA) subjected to deep learning–based image restoration (DLR) method with images subjected to hybrid iterative reconstruction (IR). We enrolled 30 patients (22 men, 8 women) who underwent coronary CTA on a 320-slice CT scanner. The images were reconstructed with hybrid IR and with DLR. The image noise in the ascending aorta, left atrium, and septal wall of the ventricle was measured on all images and the contrast-to-noise ratio (CNR) in the proximal coronary arteries was calculated. We also generated CT attenuation profiles across the proximal coronary arteries and measured the width of the edge rise distance (ERD) and the edge rise slope (ERS). Two observers visually evaluated the overall image quality using a 4-point scale (1 = poor, 4 = excellent). On DLR images, the mean image noise was lower than that on hybrid IR images (18.5 ± 2.8 HU vs. 23.0 ± 4.6 HU, p < 0.01) and the CNR was significantly higher (p < 0.01). The mean ERD was significantly shorter on DLR than on hybrid IR images, whereas the mean ERS was steeper on DLR than on hybrid IR images. The mean image quality score for hybrid IR and DLR images was 2.96 and 3.58, respectively (p < 0.01). DLR reduces the image noise and improves the image quality at coronary CTA. • Deep learning–based image restoration is a new technique that employs the deep convolutional neural network for image quality improvement. • Deep learning–based restoration reduces the image noise and improves image quality at coronary CT angiography. • This method may allow for a reduction in radiation exposure.

Precise diagnosis of intracranial hemorrhage and subtypes using a three-dimensional joint convolutional and recurrent neural network.

Abstract: To evaluate the performance of a novel three-dimensional (3D) joint convolutional and recurrent neural network (CNN-RNN) for the detection of intracranial hemorrhage (ICH) and its five subtypes (cerebral parenchymal, intraventricular, subdural, epidural, and subarachnoid) in non-contrast head CT. A total of 2836 subjects (ICH/normal, 1836/1000) from three institutions were included in this ethically approved retrospective study, with a total of 76,621 slices from non-contrast head CT scans. ICH and its five subtypes were annotated by three independent experienced radiologists, with majority voting as reference standard for both the subject level and the slice level. Ninety percent of data was used for training and validation, and the rest 10% for final evaluation. A joint CNN-RNN classification framework was proposed, with the flexibility to train when subject-level or slice-level labels are available. The predictions were compared with the interpretations from three junior radiology trainees and an additional senior radiologist. It took our algorithm less than 30 s on average to process a 3D CT scan. For the two-type classification task (predicting bleeding or not), our algorithm achieved excellent values (≥ 0.98) across all reporting metrics on the subject level. For the five-type classification task (predicting five subtypes), our algorithm achieved > 0.8 AUC across all subtypes. The performance of our algorithm was generally superior to the average performance of the junior radiology trainees for both two-type and five-type classification tasks. The proposed method was able to accurately detect ICH and its subtypes with fast speed, suggesting its potential for assisting radiologists and physicians in their clinical diagnosis workflow. • A 3D joint CNN-RNN deep learning framework was developed for ICH detection and subtype classification, which has the flexibility to train with either subject-level labels or slice-level labels. • This deep learning framework is fast and accurate at detecting ICH and its subtypes. • The performance of the automated algorithm was superior to the average performance of three junior radiology trainees in this work, suggesting its potential to reduce initial misinterpretations.

Radiomics analysis of multiparametric MRI for prediction of pathological complete response to neoadjuvant chemoradiotherapy in locally advanced rectal cancer

Abstract: To develop and validate a radiomics predictive model based on pre-treatment multiparameter magnetic resonance imaging (MRI) features and clinical features to predict a pathological complete response (pCR) in patients with locally advanced rectal cancer (LARC) after receiving neoadjuvant chemoradiotherapy (CRT). One hundred and eighty-six consecutive patients with LARC (training dataset, n = 131; validation dataset, n = 55) were enrolled in our retrospective study. A total of 1,188 imaging features were extracted from pre-CRT T2-weighted (T2-w), contrast-enhanced T1-weighted (cT1-w) and ADC images for each patient. Three steps including least absolute shrinkage and selection operator (LASSO) regression were performed to select key features and build a radiomics signature. Combining clinical risk factors, a radiomics nomogram was constructed. The predictive performance was evaluated by receiver operator characteristic (ROC) curve analysis, and then assessed with respect to its calibration, discrimination and clinical usefulness. Thirty-one of 186 patients (16.7%) achieved pCR. The radiomics signature derived from joint T2-w, ADC, and cT1-w images, comprising 12 selected features, was significantly associated with pCR status and showed better predictive performance than signatures derived from either of them alone in both datasets. The radiomics nomogram, incorporating the radiomics signature and MR-reported T-stages, also showed good discrimination, with areas under the ROC curves (AUCs) of 0.948 (95% CI, 0.907-0.989) and 0.966 (95% CI, 0.924-1.000), as well as good calibration in both datasets. Decision curve analysis confirmed its clinical usefulness. This study demonstrated that the pre-treatment radiomics nomogram can predict pCR in patients with LARC and potentially guide treatments to select patients for a “wait-and-see” policy. • Radiomics analysis of pre-CRT multiparameter MR images could predict pCR in patients with LARC. • Proposed radiomics signature from joint T2-w, ADC and cT1-w images showed better predictive performance than individual signatures. • Most of the clinical characteristics were unable to predict pCR.

Endometrial Cancer MRI staging: Updated Guidelines of the European Society of Urogenital Radiology.

Abstract: To update the 2009 ESUR endometrial cancer guidelines and propose strategies to standardize image acquisition, interpretation and reporting for endometrial cancer staging with MRI. The published evidence-based data and the opinion of experts were combined using the RAND-UCLA Appropriateness Method and formed the basis for these consensus guidelines. The responses of the experts to 81 questions regarding the details of patient preparation, MR imaging protocol, image interpretation and reporting were collected, analysed and classified as “RECOMMENDED” versus “NOT RECOMMENDED” (if at least 80% consensus among experts) or uncertain (if less than 80% consensus among experts). Consensus regarding patient preparation, MR image acquisition, interpretation and reporting was determined using the RAND-UCLA Appropriateness Method. A tailored MR imaging protocol and a standardized report were recommended. These consensus recommendations should be used as a guide for endometrial cancer staging with MRI. • MRI is recommended for initial staging of endometrial cancer. • MR imaging protocol should be tailored based on the risk of lymph node metastases. • Myometrial invasion is best assessed using combined axial-oblique T2WI, DWI and contrast-enhanced imaging. • The mnemonic “Clinical and MRI Critical TEAM” summarizes key elements of the standardized report.

Preoperative radiomics nomogram for microvascular invasion prediction in hepatocellular carcinoma using contrast-enhanced CT.

Abstract: To develop and validate a radiomics nomogram for preoperative prediction of microvascular invasion (MVI) in patients with hepatocellular carcinoma (HCC). The study included 157 patients with histologically confirmed HCC with or without MVI, and 110 patients were allocated to the training dataset and 47 to the validation dataset. Baseline clinical factor (CF) data were collected from our medical records, and radiomics features were extracted from the artery phase (AP), portal venous phase (PVP) and delay phase (DP) of preoperatively acquired CT in all patients. Radiomics analysis included tumour segmentation, feature extraction, model construction and model evaluation. A final nomogram for predicting MVI of HCC was established. Nomogram performance was assessed via both calibration and discrimination statistics. Five AP features, seven PVP features and nine DP features were effective for MVI prediction in HCC radiomics signatures. PVP radiomics signatures exhibited better performance than AP and DP radiomics signatures in the validation datasets, with the AUC 0.793. In the clinical model, age, maximum tumour diameter, alpha-fetoprotein and hepatitis B antigen were effective predictors. The final nomogram integrated the PVP radiomics signature and four CFs. Good calibration was achieved for the nomogram in both the training and validated datasets, with respective C-indexes of 0.827 and 0.820. Decision curve analysis suggested that the proposed nomogram was clinically useful, with a corresponding net benefit of 0.357. The above-described radiomics nomogram can preoperatively predict MVI in patients with HCC and may constitute a usefully clinical tool to guide subsequent personalised treatment. • No previously reported study has utilised radiomics nomograms to preoperatively predict the MVI of HCC using 3D contrast-enhanced CT imaging. • The combined radiomics clinical factor (CF) nomogram for predicting MVI achieved superior performance than either the radiomics signature or the CF nomogram alone. • Nomograms combing PVP radiomics and CF may be useful as an imaging marker for predicting MVI of HCC preoperatively and could guide personalised treatment.

Fully automated detection and segmentation of meningiomas using deep learning on routine multiparametric MRI.

Abstract: Magnetic resonance imaging (MRI) is the method of choice for imaging meningiomas. Volumetric assessment of meningiomas is highly relevant for therapy planning and monitoring. We used a multiparametric deep-learning model (DLM) on routine MRI data including images from diverse referring institutions to investigate DLM performance in automated detection and segmentation of meningiomas in comparison to manual segmentations. We included 56 of 136 consecutive preoperative MRI datasets [T1/T2-weighted, T1-weighted contrast-enhanced (T1CE), FLAIR] of meningiomas that were treated surgically at the University Hospital Cologne and graded histologically as tumour grade I (n = 38) or grade II (n = 18). The DLM was trained on an independent dataset of 249 glioma cases and segmented different tumour classes as defined in the brain tumour image segmentation benchmark (BRATS benchmark). The DLM was based on the DeepMedic architecture. Results were compared to manual segmentations by two radiologists in a consensus reading in FLAIR and T1CE. The DLM detected meningiomas in 55 of 56 cases. Further, automated segmentations correlated strongly with manual segmentations: average Dice coefficients were 0.81 ± 0.10 (range, 0.46-0.93) for the total tumour volume (union of tumour volume in FLAIR and T1CE) and 0.78 ± 0.19 (range, 0.27-0.95) for contrast-enhancing tumour volume in T1CE. The DLM yielded accurate automated detection and segmentation of meningioma tissue despite diverse scanner data and thereby may improve and facilitate therapy planning as well as monitoring of this highly frequent tumour entity. • Deep learning allows for accurate meningioma detection and segmentation • Deep learning helps clinicians to assess patients with meningiomas • Meningioma monitoring and treatment planning can be improved

Preoperative prediction of microvascular invasion in hepatocellular cancer: a radiomics model using Gd-EOB-DTPA-enhanced MRI

Abstract: Preoperative prediction of microvascular invasion (MVI) in patients with hepatocellular cancer (HCC) is important for surgery strategy making. We aimed to develop and validate a combined intratumoural and peritumoural radiomics model based on gadolinium-ethoxybenzyl-diethylenetriamine (Gd-EOB-DTPA)-enhanced magnetic resonance imaging (MRI) for preoperative prediction of MVI in primary HCC patients. This study included a training cohort of 110 HCC patients and a validating cohort of 50 HCC patients. All the patients underwent preoperative Gd-EOB-DTPA-enhanced MRI examination and curative hepatectomy. The volumes of interest (VOIs) around the hepatic lesions including intratumoural and peritumoural regions were manually delineated in the hepatobiliary phase of MRI images, from which quantitative features were extracted and analysed. In the training cohort, machine-learning method was applied for dimensionality reduction and selection of the extracted features. The proportion of MVI-positive patients was 38.2% and 40.0% in the training and validation cohort, respectively. Supervised machine learning selected ten features to establish a predictive model for MVI. The area under the receiver operating characteristic curve (AUC), sensitivity, specificity of the combined intratumoural and peritumoural radiomics model in the training and validation cohort were 0.85 (95% confidence interval (CI), 0.77–0.93), 88.2%, 76.2%, and 0.83 (95% CI, 0.71–0.95), 90.0%, 75.0%, respectively. We evaluate quantitative Gd-EOB-DTPA-enhanced MRI image features of both intratumoural and peritumoural regions and provide an effective radiomics-based model for the prediction of MVI in HCC patients, and may therefore help clinicians make precise decisions regarding treatment before the surgery. • An effective radiomics model for prediction of microvascular invasion in HCC patients is established. • The radiomics model is superior to the radiologist in prediction of MVI. • The radiomics model can help clinicians in pretreatment decision making.

Inter-observer variability of manual contour delineation of structures in CT.

Abstract: To quantify the inter-observer variability of manual delineation of lesions and organ contours in CT to establish a reference standard for volumetric measurements for clinical decision making and for the evaluation of automatic segmentation algorithms Eleven radiologists manually delineated 3193 contours of liver tumours (896), lung tumours (1085), kidney contours (434) and brain hematomas (497) on 490 slices of clinical CT scans A comparative analysis of the delineations was then performed to quantify the inter-observer delineation variability with standard volume metrics and with new group-wise metrics for delineations produced by groups of observers The mean volume overlap variability values and ranges (in %) between the delineations of two observers were: liver tumours 178 [-58,+72]%, lung tumours 208 [-88,+102]%, kidney contours 88 [-08,+12]% and brain hematomas 18 [-60,+60] % For any two randomly selected observers, the mean delineation volume overlap variability was 5–57% The mean variability captured by groups of two, three and five observers was 37%, 53% and 72%; eight observers accounted for 75–94% of the total variability For all cases, 385% of the delineation non-agreement was due to parts of the delineation of a single observer disagreeing with the others No statistical difference was found for the delineation variability between the observers based on their expertise The variability in manual delineations for different structures and observers is large and spans a wide range across a variety of structures and pathologies Two and even three observers may not be sufficient to establish the full range of inter-observer variability • This study quantifies the inter-observer variability of manual delineation of lesions and organ contours in CT • The variability of manual delineations between two observers can be significant Two and even three observers capture only a fraction of the full range of inter-observer variability observed in common practice • Inter-observer manual delineation variability is necessary to establish a reference standard for radiologist training and evaluation and for the evaluation of automatic segmentation algorithms

Ultrasound-based radiomics score: a potential biomarker for the prediction of microvascular invasion in hepatocellular carcinoma

Abstract: To develop an ultrasound (US)-based radiomics score for preoperative prediction of microvascular invasion (MVI) in hepatocellular carcinoma (HCC). Between January 1, 2012, and October 31, 2017, a total of 482 HCC patients who underwent contrast-enhanced ultrasound (CEUS) were retrospectively reviewed. The study population was divided into a training cohort (n = 341) and a validation cohort (n = 141) based on a cutoff time of January 1, 2016. Radiomics features were extracted from the grayscale US images of HCC. After features selection, a radiomics score was developed from the training cohort. The incremental value of the radiomics score to the clinic-pathological factors for MVI prediction was assessed in the validation cohort with respect to discrimination, calibration, and clinical usefulness. The US-based radiomics score consisted of six selected features. Multivariate logistic regression analysis showed that the radiomics score, alpha-fetoprotein (AFP), and tumor size were independent predictors of MVI. The radiomics nomogram (based on the three factors) showed better performance for MVI detection (area under the curve [AUC] 0.731[0.647, 0.815] than the clinical nomogram (based on AFP and tumor size) (0.634 [0.543, 0.724]) (p = 0.015). Both nomograms showed good calibration. Decision curve analysis demonstrated that in terms of clinical usefulness, the radiomics nomogram outperformed the clinical nomogram. The US-based radiomics score was an independent predictor of MVI in HCC. Combining the radiomics score with clinical factors improved the prediction efficacy. • Radiomics can be applied in US images. • US-based radiomics score was an independent predictor of MVI. • Radiomics nomogram incorporated with the radiomics score showed good performance for MVI prediction.

Diagnostic value of MRI-PDFF for hepatic steatosis in patients with non-alcoholic fatty liver disease: a meta-analysis

Abstract: To systematically review studies about the diagnostic accuracy of magnetic resonance imaging proton density fat fraction (MRI-PDFF) in the classification of hepatic steatosis grade in patients with non-alcoholic fatty liver disease (NAFLD). Areas under the summary receiver operating characteristic curves (AUROC), sensitivity, specificity, overall diagnostic odds ratio (DOR), diagnostic score, positive likelihood ratio (+LR), and negative likelihood ratio (−LR) for MRI-PDFF in classification of steatosis grades 0 vs. 1–3, 0–1 vs. 2–3, and 0–2 vs. 3 were compared and analyzed. A total of 6 studies were included in this meta-analysis (n = 635). The summary AUROC values of MRI-PDFF for classifying steatosis grades 0 vs. 1–3, 0–1 vs. 2–3, and 0–2 vs. 3 were 0.98, 0.91, and 0.90, respectively. Pooled sensitivity and specificity of MRI-PDFF for classifying steatosis grades 0 vs. 1–3, 0–1 vs. 2–3, and 0–2 vs. 3 were 0.93 and 0.94, 0.74 and 0.90, and 0.74 and 0.87, respectively. Summary +LR and −LR of MRI-PDFF for classifying steatosis grades 0 vs. 1–3, 0–1 vs. 2–3, and 0–2 vs. 3 were 16.21 (95%CI, 4.72–55.67) and 0.08 (95%CI, 0.04–0.15), 7.19 (95%CI, 5.04–10.26) and 0.29 (95%CI, 0.22–0.38), and 5.89 (95%CI, 4.27–8.13) and 0.29 (95%CI, 0.21–0.41), respectively. Our meta-analysis suggests that MRI-PDFF has excellent diagnostic value for assessment of hepatic fat content and classification of histologic steatosis in patients with NAFLD. • MRI-PDFF has significant diagnostic value for hepatic steatosis in patients with NAFLD. • MRI-PDFF may be used to classify grade of hepatic steatosis with high sensitivity and specificity.

Can we reduce the workload of mammographic screening by automatic identification of normal exams with artificial intelligence? A feasibility study

Abstract: Purpose To study the feasibility of automatically identifying normal digital mammography (DM) exams with artificial intelligence (AI) to reduce the breast cancer screening reading workload.

Radiomics signature: a biomarker for the preoperative discrimination of lung invasive adenocarcinoma manifesting as a ground-glass nodule

Abstract: To identify the radiomics signature allowing preoperative discrimination of lung invasive adenocarcinomas from non-invasive lesions manifesting as ground-glass nodules. This retrospective primary cohort study included 160 pathologically confirmed lung adenocarcinomas. Radiomics features were extracted from preoperative non-contrast CT images to build a radiomics signature. The predictive performance and calibration of the radiomics signature were evaluated using intra-cross (n=76), external non-contrast-enhanced CT (n=75) and contrast-enhanced CT (n=84) validation cohorts. The performance of radiomics signature and CT morphological and quantitative indices were compared. 355 three-dimensional radiomics features were extracted, and two features were identified as the best discriminators to build a radiomics signature. The radiomics signature showed a good ability to discriminate between invasive adenocarcinomas and non-invasive lesions with an accuracy of 86.3%, 90.8%, 84.0% and 88.1%, respectively, in the primary and validation cohorts. It remained an independent predictor after adjusting for traditional preoperative factors (odds ratio 1.87, p < 0.001) and demonstrated good calibration in all cohorts. It was a better independent predictor than CT morphology or mean CT value. The radiomics signature showed good predictive performance in discriminating between invasive adenocarcinomas and non-invasive lesions. Being a non-invasive biomarker, it could assist in determining therapeutic strategies for lung adenocarcinoma. • The radiomics signature was a non-invasive biomarker of lung invasive adenocarcinoma. • The radiomics signature outweighed CT morphological and quantitative indices. • A three-centre study showed that radiomics signature had good predictive performance.

Radiomic nomogram for prediction of axillary lymph node metastasis in breast cancer.

Abstract: To develop a radiomic nomogram for preoperative prediction of axillary lymph node (LN) metastasis in breast cancer patients. Preoperative magnetic resonance imaging data from 411 breast cancer patients was studied. Patients were assigned to either a training cohort (n = 279) or a validation cohort (n = 132). Eight hundred eight radiomic features were extracted from the first phase of T1-DCE images. A support vector machine was used to develop a radiomic signature, and logistic regression was used to develop a nomogram. The radiomic signature based on 12 LN status–related features was constructed to predict LN metastasis, its prediction ability was moderate, with an area under the curve (AUC) of 0.76 and 0.78 in training and validation cohorts, respectively. Based on a radiomic signature and clinical features, a nomogram was developed and showed excellent predictive ability for LN metastasis (AUC 0.84 and 0.87 in training and validation sets, respectively). Another radiomic signature was constructed to distinguish the number of metastatic LNs (less than 2 positive nodes/more than 2 positive nodes), which also showed moderate performance (AUC 0.79). We developed a nomogram and a radiomic signature that can be used to identify LN metastasis and distinguish the number of metastatic LNs (less than 2 positive nodes/more than 2 positive nodes). Both nomogram and radiomic signature can be used as tools to assist clinicians in assessing LN metastasis in breast cancer patients. • ALNM is an important factor affecting breast cancer patients’ treatment and prognosis. • Traditional imaging examinations have limited value for evaluating axillary LNs status. • We developed a radiomic nomogram based on MR imagings to predict LN metastasis.

Radiomics and machine learning may accurately predict the grade and histological subtype in meningiomas using conventional and diffusion tensor imaging

Abstract: Preoperative, noninvasive prediction of the meningioma grade is important because it influences the treatment strategy. The purpose of this study was to evaluate the role of radiomics features of postcontrast T1-weighted images (T1C), apparent diffusion coefficient (ADC), and fractional anisotropy (FA) maps, based on the entire tumor volume, in the differentiation of grades and histological subtypes of meningiomas. One hundred thirty-six patients with pathologically diagnosed meningiomas (108 low-grade [benign], 28 high-grade [atypical and anaplastic]), who underwent T1C and diffusion tensor imaging, were included in the discovery set. The T1C image, ADC, and FA maps were analyzed to derive volume-based data of the entire tumor. Radiomics features were correlated with meningioma grades and histological subtypes. Various machine learning classifiers were trained to build classification models to predict meningioma grades. We tested the model in a validation set (58 patients; 46 low-grade; 12 high-grade). The machine learning classifiers showed variable performances depending on the machine learning algorithms. The best classification system for the prediction of meningioma grades had an area under the curve of 0.86 (95% confidence interval [CI], 0.74–0.98) in the validation set. The accuracy, sensitivity, and specificity of the best classifier were 89.7, 75.0, and 93.5% in the validation set, respectively. Various texture parameters differed significantly between fibroblastic and non-fibroblastic subtypes. Radiomics feature-based machine learning classifiers of T1C images, ADC, and FA maps are useful for differentiating meningioma grades. • Preoperative, noninvasive differentiation of the meningioma grade is important because it influences the treatment strategy. • Radiomics feature-based machine learning classifiers of T1C images, ADC, and FA maps are useful for differentiating meningioma grades. • In benign meningiomas, there were significant differences in the various texture parameters between fibroblastic and non-fibroblastic meningioma subtypes.

Predicting the grade of hepatocellular carcinoma based on non-contrast-enhanced MRI radiomics signature

Abstract: This study was conducted in order to investigate the value of magnetic resonance imaging (MRI)-based radiomics signatures for the preoperative prediction of hepatocellular carcinoma (HCC) grade. Data from 170 patients confirmed to have HCC by surgical pathology were divided into a training group (n = 125) and a test group (n = 45). The radiomics features of tumours based on both T1-weighted imaging (WI) and T2WI were extracted by using Matrix Laboratory (MATLAB), and radiomics signatures were generated using the least absolute shrinkage and selection operator (LASSO) logistic regression model. The predicted values of pathological HCC grades using radiomics signatures, clinical factors (including age, sex, tumour size, alpha fetoprotein (AFP) level, history of hepatitis B, hepatocirrhosis, portal vein tumour thrombosis, portal hypertension and pseudocapsule) and the combined models were assessed. Radiomics signatures could successfully categorise high-grade and low-grade HCC cases (p < 0.05) in both the training and test datasets. Regarding the performances of clinical factors, radiomics signatures and the combined clinical and radiomics signature (from the combined T1WI and T2WI images) models for HCC grading prediction, the areas under the curve (AUCs) were 0.600, 0.742 and 0.800 in the test datasets, respectively. Both the AFP level and radiomics signature were independent predictors of HCC grade (p < 0.05). Radiomics signatures may be important for discriminating high-grade and low-grade HCC cases. The combination of the radiomics signatures with clinical factors may be helpful for the preoperative prediction of HCC grade. • The radiomics signature based on non-contrast-enhanced MR images was significantly associated with the pathological grade of HCC. • The radiomics signatures based on T1WI or T2WI images performed similarly at predicting the pathological grade of HCC. • Combining the radiomics signature and clinical factors (including age, sex, tumour size, AFP level, history of hepatitis B, hepatocirrhosis, portal vein tumour thrombosis, portal hypertension and pseudocapsule) may be helpful for the preoperative prediction of HCC grade.

Identifying EGFR mutations in lung adenocarcinoma by noninvasive imaging using radiomics features and random forest modeling

Abstract: The tyrosine kinase inhibitor (TKI)-sensitive mutations of the epidermal growth factor receptor (EGFR) gene is essential in the treatment of lung adenocarcinoma. To overcome the difficulty of EGFR gene test in situations where surgery and biopsy samples are too risky to obtain, we tried a noninvasive imaging method using radiomics features and random forest models. Five hundred three lung adenocarcinoma patients who received surgery-based treatment were included in this study. The diagnosis and EGFR gene test were based on resections. TKI-sensitive mutations were found in 60.8% of the patients. CT scans before any invasive operation were gathered and analyzed to extract quantitative radiomics features and build random forest classifiers to identify EGFR mutants from wild types. Clinical features (sex and smoking history) were added to the image-based model. The model was trained on a set of 345 patients and validated on an independent test group (n = 158) using the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. The performance of the random forest model with 94 radiomics features reached an AUC of 0.802. Its AUC was further improved to 0.828 by adding sex and smoking history. The sensitivity and specificity are 60.6% and 85.1% at the best diagnostic decision point. Our results showed that radiomics could not only reflect the genetic differences among tumors but also have diagnostic value and the potential to be a diagnostic tool. • Radiomics provides a potential noninvasive method for the prediction of EGFR mutation status. • In situations where surgeries and biopsy are not available, CT image-based radiomics models could help to make treatment decisions. • The accuracy, sensitivity, and specificity still need to be improved before the image-based EGFR identifier could be used in clinics.

CT texture analysis of pancreatic cancer

Abstract: We investigated the value of CT texture analysis (CTTA) in predicting prognosis of unresectable pancreatic cancer. Sixty patients with unresectable pancreatic cancers at presentation were enrolled for post-processing with CTTA using commercially available software (TexRAD Ltd, Cambridge, UK). The largest cross-section of the tumour on axial CT was chosen to draw a region-of-interest. CTTA parameters (mean value of positive pixels (MPP), kurtosis, entropy, skewness), arterial and venous invasion, metastatic disease and tumour size were correlated with overall and progression-free survivals. The median overall and progression-free survivals of cohort were 13.3 and 7.8 months, respectively. On multivariate Cox proportional hazard regression analysis, presence of metastatic disease at presentation had the highest association with overall survival (p = 0.003–0.05) and progression-free survival (p < 0.001 to p = 0.004). MPP at medium spatial filter was significantly associated with poor overall survival (p = 0.04). On Kaplan–Meier survival analysis of CTTA parameters at medium spatial filter, MPP of more than 31.625 and kurtosis of more than 0.565 had significantly worse overall survival (p = 0.036 and 0.028, respectively). CTTA features were significantly associated with overall survival in pancreas cancer, particularly in patients with non-metastatic, locally advanced disease. • CT texture analysis is easy to perform on contrast-enhanced CT. • CT texture analysis can determine prognosis in patients with unresectable pancreas cancer. • The best predictors of poor prognosis were high kurtosis and MPP.

Clear Cell Renal Cell Carcinoma: Machine Learning-Based Quantitative Computed Tomography Texture Analysis for Prediction of Fuhrman Nuclear Grade

Abstract: To evaluate the performance of quantitative computed tomography (CT) texture analysis using different machine learning (ML) classifiers for discriminating low and high nuclear grade clear cell renal cell carcinomas (cc-RCCs) This retrospective study included 53 patients with pathologically proven 54 cc-RCCs (31 low-grade [grade 1 or 2]; 23 high-grade [grade 3 or 4]) In one patient, two synchronous cc-RCCs were included in the analysis Mean age was 575 years Thirty-four (641%) patients were male and 19 were female (359%) Mean tumour size based on the maximum diameter was 574 mm (range, 16–145 mm) Forty patients underwent radical nephrectomy and 13 underwent partial nephrectomy Following pre-processing steps, two-dimensional CT texture features were extracted using portal-phase contrast-enhanced CT Reproducibility of texture features was assessed with the intra-class correlation coefficient (ICC) Nested cross-validation with a wrapper-based algorithm was used in feature selection and model optimisation The ML classifiers were support vector machine (SVM), multilayer perceptron (MLP, a sort of neural network), naive Bayes, k-nearest neighbours, and random forest The performance of the classifiers was compared by certain metrics Among 279 texture features, 241 features with an ICC equal to or higher than 080 (excellent reproducibility) were included in the further feature selection process The best model was created using SVM The selected subset of features for SVM included five co-occurrence matrix (ICC range, 0885–0998), three run-length matrix (ICC range, 0889–0992), one gradient (ICC = 0998), and four Haar wavelet features (ICC range, 0941–0997) The overall accuracy, sensitivity (for detecting high-grade cc-RCCs), specificity (for detecting high-grade cc-RCCs), and overall area under the curve of the best model were 851%, 913%, 806%, and 0860, respectively The ML-based CT texture analysis can be a useful and promising non-invasive method for prediction of low and high Fuhrman nuclear grade cc-RCCs • Based on the percutaneous biopsy literature, ML-based CT texture analysis has a comparable predictive performance with percutaneous biopsy • Highest predictive performance was obtained with use of the SVM • SVM correctly classified 851% of cc-RCCs in terms of nuclear grade, with an AUC of 0860

Deep-learning-based detection and segmentation of organs at risk in nasopharyngeal carcinoma computed tomographic images for radiotherapy planning

Abstract: Accurate detection and segmentation of organs at risks (OARs) in CT image is the key step for efficient planning of radiation therapy for nasopharyngeal carcinoma (NPC) treatment. We develop a fully automated deep-learning-based method (termed organs-at-risk detection and segmentation network (ODS net)) on CT images and investigate ODS net performance in automated detection and segmentation of OARs. The ODS net consists of two convolutional neural networks (CNNs). The first CNN proposes organ bounding boxes along with their scores, and then a second CNN utilizes the proposed bounding boxes to predict segmentation masks for each organ. A total of 185 subjects were included in this study for statistical comparison. Sensitivity and specificity were performed to determine the performance of the detection and the Dice coefficient was used to quantitatively measure the overlap between automated segmentation results and manual segmentation. Paired samples t tests and analysis of variance were employed for statistical analysis. ODS net provides an accurate detection result with a sensitivity of 0.997 to 1 for most organs and a specificity of 0.983 to 0.999. Furthermore, segmentation results from ODS net correlated strongly with manual segmentation with a Dice coefficient of more than 0.85 in most organs. A significantly higher Dice coefficient for all organs together (p = 0.0003 < 0.01) was obtained in ODS net (0.861 ± 0.07) than in fully convolutional neural network (FCN) (0.8 ± 0.07). The Dice coefficients of each OAR did not differ significantly between different T-staging patients. The ODS net yielded accurate automated detection and segmentation of OARs in CT images and thereby may improve and facilitate radiotherapy planning for NPC. • A fully automated deep-learning method (ODS net) is developed to detect and segment OARs in clinical CT images. • This deep-learning-based framework produces reliable detection and segmentation results and thus can be useful in delineating OARs in NPC radiotherapy planning. • This deep-learning-based framework delineating a single image requires approximately 30 s, which is suitable for clinical workflows.

Multiparametric MRI of the bladder: inter-observer agreement and accuracy with the Vesical Imaging-Reporting and Data System (VI-RADS) at a single reference center

Abstract: To evaluate accuracy and inter-observer variability using Vesical Imaging-Reporting and Data System (VI-RADS) for discrimination between non-muscle invasive bladder cancer (NMIBC) and muscle-invasive bladder cancer (MIBC). Between September 2017 and July 2018, 78 patients referred for suspected bladder cancer underwent multiparametric MRI of the bladder (mpMRI) prior to transurethral resection of bladder tumor (TURBT). All mpMRI were reviewed by two radiologists, who scored each lesion according to VI-RADS. Sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated for each VI-RADS cutoff. Receiver operating characteristics curves were used to evaluate the performance of mpMRI. The Ƙ statistics was used to estimate inter-reader agreement. Seventy-five patients were included in the final analysis, 53 with NMIBC and 22 with MIBC. Sensitivity and specificity were 91% and 89% for reader 1 and 82% and 85% for reader 2 respectively when the cutoff VI-RADS > 2 was used to define MIBC. At the same cutoff, PPV and NPV were 77% and 96% for reader 1 and 69% and 92% for reader 2. When the cutoff VI-RADS > 3 was used, sensitivity and specificity were 82% and 94% for reader 1 and 77% and 89% for reader 2. Corresponding PPV and NPV were 86% and 93% for reader 1 and 74% and 91% for reader 2. Area under curve was 0.926 and 0.873 for reader 1 and 2 respectively. Inter-reader agreement was good for the overall score (Ƙ = 0.731). VI-RADS is accurate in differentiating MIBC from NMIBC. Inter-reader agreement is overall good. • Traditionally, the local staging of bladder cancer relies on transurethral resection of bladder tumor. • However, transurethral resection of bladder tumor carries a significant risk of understaging a cancer; therefore, more accurate, faster, and non-invasive staging techniques are needed to improve outcomes. • Multiparametric MRI has proved to be the best imaging modality for local staging; therefore, its use in suitable patients has the potential to expedite radical treatment when necessary and non-invasive diagnosis in patients with poor fitness.

Breast density implications and supplemental screening

Abstract: Digital breast tomosynthesis (DBT) has been widely implemented in place of 2D mammography, although it is less effective in women with extremely dense breasts. Breast ultrasound detects additional early-stage, invasive breast cancers when combined with mammography; however, its relevant limitations, including the shortage of trained operators, operator dependence and small field of view, have limited its widespread implementation. Automated breast sonography (ABS) is a promising technique but the time to interpret and false-positive rates need to be improved. Supplemental screening with contrast-enhanced magnetic resonance imaging (MRI) in high-risk women reduces late-stage disease; abbreviated MRI protocols may reduce cost and increase accessibility to women of average risk with dense breasts. Contrast-enhanced digital mammography (CEDM) and molecular breast imaging improve cancer detection but require further validation for screening and direct biopsy guidance should be implemented for any screening modality. This article reviews the status of screening women with dense breasts. KEY POINTS: • The sensitivity of mammography is reduced in women with dense breasts. Supplemental screening with US detects early-stage, invasive breast cancers. • Tomosynthesis reduces recall rate and increases cancer detection rate but is less effective in women with extremely dense breasts. • Screening MRI improves early diagnosis of breast cancer more than ultrasound and is currently recommended for women at high risk. Risk assessment is needed, to include breast density, to ascertain who should start early annual MRI screening.

Deep learning for liver tumor diagnosis part II: convolutional neural network interpretation using radiologic imaging features

Abstract: To develop a proof-of-concept “interpretable” deep learning prototype that justifies aspects of its predictions from a pre-trained hepatic lesion classifier. A convolutional neural network (CNN) was engineered and trained to classify six hepatic tumor entities using 494 lesions on multi-phasic MRI, described in Part 1. A subset of each lesion class was labeled with up to four key imaging features per lesion. A post hoc algorithm inferred the presence of these features in a test set of 60 lesions by analyzing activation patterns of the pre-trained CNN model. Feature maps were generated that highlight regions in the original image that correspond to particular features. Additionally, relevance scores were assigned to each identified feature, denoting the relative contribution of a feature to the predicted lesion classification. The interpretable deep learning system achieved 76.5% positive predictive value and 82.9% sensitivity in identifying the correct radiological features present in each test lesion. The model misclassified 12% of lesions. Incorrect features were found more often in misclassified lesions than correctly identified lesions (60.4% vs. 85.6%). Feature maps were consistent with original image voxels contributing to each imaging feature. Feature relevance scores tended to reflect the most prominent imaging criteria for each class. This interpretable deep learning system demonstrates proof of principle for illuminating portions of a pre-trained deep neural network’s decision-making, by analyzing inner layers and automatically describing features contributing to predictions. • An interpretable deep learning system prototype can explain aspects of its decision-making by identifying relevant imaging features and showing where these features are found on an image, facilitating clinical translation. • By providing feedback on the importance of various radiological features in performing differential diagnosis, interpretable deep learning systems have the potential to interface with standardized reporting systems such as LI-RADS, validating ancillary features and improving clinical practicality. • An interpretable deep learning system could potentially add quantitative data to radiologic reports and serve radiologists with evidence-based decision support.

Development of a radiomics nomogram based on the 2D and 3D CT features to predict the survival of non-small cell lung cancer patients.

Abstract: The aim of this study was to develop a radiomics nomogram by combining the optimized radiomics signatures extracted from 2D and/or 3D CT images and clinical predictors to assess the overall survival of patients with non-small cell lung cancer (NSCLC). One training cohort of 239 and two validation datasets of 80 and 52 NSCLC patients were enrolled in this study. Nine hundred seventy-five radiomics features were extracted from each patient’s 2D and 3D CT images. Least absolute shrinkage and selection operator (LASSO) regression was used to select features and generate a radiomics signature. Cox hazard survival analysis and Kaplan-Meier were performed in both cohorts. The radiomics nomogram was developed by integrating the optimized radiomics signature and clinical predictors, its calibration and discrimination were evaluated. The radiomics signatures were significantly associated with NSCLC patients’ survival time. The signature derived from the combined 2D and 3D features showed a better prognostic performance than those from 2D or 3D alone. Our radiomics nomogram integrated the optimal radiomics signature with clinical predictors showed a significant improvement in the prediction of patients’ survival compared with clinical predictors alone in the validation cohort. The calibration curve showed predicted survival time was very close to the actual one. The radiomics signature from the combined 2D and 3D features further improved the predicted accuracy of survival prognosis for the patients with NSCLC. Combination of the optimal radiomics signature and clinical predictors performed better for individualied survival prognosis estimation in patients with NSCLC. These findings might affect trearment strategies and enable a step forward for precise medicine. • We found both 2D and 3D radiomics signature have favorable prognosis, but 3D signature had a better performance. • The radiomics signature generated from the combined 2D and 3D features had a better predictive performance than those from 2D or 3D features. • Integrating the optimal radiomics signature with clinical predictors significantly improved the predictive power in patients’ survival compared with clinical TNM staging alone.

Pretreatment prediction of immunoscore in hepatocellular cancer: a radiomics-based clinical model based on Gd-EOB-DTPA-enhanced MRI imaging

Abstract: Immunoscore evaluates the density of CD3+ and CD8+ T cells in both the tumor core and invasive margin. Pretreatment prediction of immunoscore in hepatocellular cancer (HCC) is important for precision immunotherapy. We aimed to develop a radiomics model based on gadolinium-ethoxybenzyl-diethylenetriamine (Gd-EOB-DTPA)-enhanced MRI for pretreatment prediction of immunoscore (0–2 vs. 3–4) in HCC. The study included 207 (training cohort: n = 150; validation cohort: n = 57) HCC patients with hepatectomy who underwent preoperative Gd-EOB-DTPA-enhanced MRI. The volumes of interest enclosing hepatic lesions including intratumoral and peritumoral regions were manually delineated in the hepatobiliary phase of MRI images, from which 1044 quantitative features were extracted and analyzed. Extremely randomized tree method was used to select radiomics features for building radiomics model. Predicting performance in immunoscore was compared among three models: (1) using only intratumoral radiomics features (intratumoral radiomics model); (2) using combined intratumoral and peritumoral radiomics features (combined radiomics model); (3) using clinical data and selected combined radiomics features (combined radiomics-based clinical model). The combined radiomics model showed a better predicting performance in immunoscore than intratumoral radiomics model (AUC, 0.904 (95% CI 0.855–0.953) vs. 0.823 (95% CI 0.747–0.899)). The combined radiomics-based clinical model showed an improvement over the combined radiomics model in predicting immunoscore (AUC, 0·926 (95% CI 0·884–0·967) vs. 0·904 (95% CI 0·855–0·953)), although differences were not statistically significant. Results were confirmed in validation cohort and calibration curves showed good agreement. The MRI-based combined radiomics nomogram is effective in predicting immunoscore in HCC and may help making treatment decisions. • Radiomics obtained from Gd-EOB-DTPA-enhanced MRI help predicting immunoscore in hepatocellular carcinoma. • Combined intratumoral and peritumoral radiomics are superior to intratumoral radiomics only in predicting immunoscore. • We developed a combined clinical and radiomicsnomogram to predict immunoscore in hepatocellular carcinoma.

Comparison of radiomics machine-learning classifiers and feature selection for differentiation of sacral chordoma and sacral giant cell tumour based on 3D computed tomography features

Abstract: We aimed to identify optimal machine-learning methods for preoperative differentiation of sacral chordoma (SC) and sacral giant cell tumour (SGCT) based on 3D non-enhanced computed tomography (CT) and CT-enhanced (CTE) features. A total of 95 patients were divided into a training set and a validation set. Three best feature selection methods (Relief, least absolute shrinkage and selection operator (LASSO) and Random Forest (RF)) and three classification methods, including generalised linear models (GLM), support vector machines (SVM) and RF, were compared for their performance in distinguishing SC and SGCT. The performance of the radiomics model was investigated via area under the receiver-operating characteristic curve (AUC) and accuracy (ACC) analysis. The selection method LASSO + classifier GLM had the highest AUC of 0.984 and ACC of 0.897 in the validating set, followed by Relief + GLM (AUC = 0.909, ACC = 0.862) and LASSO + SVM (AUC = 0.900, ACC = 0.862) based on CTE features. For CT features, RF + GLM had the highest AUC of 0.889, while LASSO + GLM achieved a high ACC of 0.793 in the validating set. Regardless of the methods, CTE features significantly outperformed those from CT for the differentiation of SC and SGCT (ZAUC = -3.029, ZACC = -4.553; p < 0.05). Our study demonstrated CTE features performed better than CT features. The selection method LASSO + classifier GLM had the best performance in differentiation of SC and SGCT, which could enhance the application of radiomics methods in sacral tumours. • Sacral chordoma and sacral giant cell tumour are the two most common primary tumours of the sacrum with many common clinical and imaging characteristics. • A radiomics model helps clinicians to identify the histology of a sacral tumour. • CTE features should be preferred.