Xuedong Deng

Bio: Xuedong Deng is an academic researcher from Nanjing Medical University. The author has an hindex of 1, co-authored 2 publications receiving 1 citations.

Ultrasonographic Segmentation of Fetal Lung with Deep Learning

Abstract: The morbidity and mortality of the fetus is related closely with the neonatal respiratory morbidity, which was caused by the immaturity of the fetal lung primarily. The amniocentesis has been used in clinics to evaluate the maturity of the fetal lung, which is invasive, expensive and time-consuming. Ultrasonography has been developed to examine the fetal lung quantitatively in the past decades as a non-invasive method. However, the contour of the fetal lung required by existing studies was delineated in manual. An automated segmentation approach could not only improve the objectiveness of those studies, but also offer a quantitative way to monitor the development of the fetal lung in terms of morphological parameters based on the segmentation. In view of this, we proposed a deep learning model for automated fetal lung segmentation and measurement. The model was constructed based on the U-Net. It was trained by 3500 data sets augmented from 250 ultrasound images with both the fetal lung and heart manually delineated, and then tested on 50 ultrasound data sets. With the proposed method, the fetal lung and cardiac area were automatically segmented with the accuracy, average IoU, sensitivity and precision being 0.98, 0.79, 0.881 and 0.886, respectively.

Performance Analysis of Squeezenet and Densenet on Fetal Brain MRI Dataset

Abstract: Monitoring fetal growth is an essential activity throughout gestation period. It is necessary for the well-being of the mother and the child so that any abnormalities can be diagnosed at the earlier stage itself and preventive measures can be taken. MRI Image analysis is necessary to identify the anomalies that have a higher probability of occurrence during the first and second trimesters. Fetal brain MRI images are obtained and analysed for any disorder in the neurodevelopment of the embryo. Classification of the fetal brain image into normal and abnormal cases is done using deep learning algorithms like densenet121 and squeezenet1_0. Densenet produced an accuracy of 98.7% when compared to squeezenet whose accuracy is 96.3%. The performance of densenet is high in case of various metrics like precision, recall, sensitivity and F1 Score.

Xception-ResNet Autoencoder for Pneumothorax Segmentation

Abstract: Computer vision has made a significant advance in the medical imaging. Pneumothorax is a severe ailment that can be fatal if the patient does not receive proper care. The main way for diagnosing Pneumothorax is by examining the Chest X-ray by a specialist. The urge of experienced radiologists to anticipate whether someone is suffering from pneumothorax or not by examining chest X-rays is indisputable. Such a facility is not available to everyone. Furthermore, in some circumstances, quick diagnosis is required. In this paper We present a deep learning based image segmentation model capable of predicting Pneumothorax cases by localizing it in chest x-ray exam to help the doctor in making this critical choice. Deep Learning has demonstrated its value in multiple domains, outperforming several state-of-the-arts methods. We seek to overcome this challenge by leveraging deep learning capabilities. We used U-Net architecture with Xception as the encoder and ResNet as a decoder. We obtained encouraging findings, and U-Net works exceptionally well in medical imaging. Our work is listed with in as semantic segmentation. With 77.8 (±0.15), our technique obtains a good outcome in terms of Intersection over Union.

Integrating multi-type aberrations from DNA and RNA through dynamic mapping gene space for subtype-specific breast cancer driver discovery

Abstract: Driver event discovery is a crucial demand for breast cancer diagnosis and therapy. In particular, discovering subtype-specificity of drivers can prompt the personalized biomarker discovery and precision treatment of cancer patients. Still, most of the existing computational driver discovery studies mainly exploit the information from DNA aberrations and gene interactions. Notably, cancer driver events would occur due to not only DNA aberrations but also RNA alternations, but integrating multi-type aberrations from both DNA and RNA is still a challenging task for breast cancer drivers. On the one hand, the data formats of different aberration types also differ from each other, known as data format incompatibility. On the other hand, different types of aberrations demonstrate distinct patterns across samples, known as aberration type heterogeneity. To promote the integrated analysis of subtype-specific breast cancer drivers, we design a “splicing-and-fusing” framework to address the issues of data format incompatibility and aberration type heterogeneity simultaneously. To overcome the data format incompatibility, the “splicing-step” employs a knowledge graph structure to connect multi-type aberrations from the DNA and RNA data into a unified formation. To tackle the aberration type heterogeneity, the “fusing-step” adopts a dynamic mapping gene space integration approach to represent the multi-type information by vectorized profiles. The experiments also demonstrate the advantages of our approach in both the integration of multi-type aberrations from DNA and RNA and the discovery of subtype-specific breast cancer drivers. In summary, our “splicing-and-fusing” framework with knowledge graph connection and dynamic mapping gene space fusion of multi-type aberrations data from DNA and RNA can successfully discover potential breast cancer drivers with subtype-specificity indication.

Transfer Learning U-Net Deep Learning for Lung Ultrasound Segmentation.

Abstract: Transfer learning (TL) for medical image segmentation helps deep learning models achieve more accurate performances when there are scarce medical images. This study focuses on completing segmentation of the ribs from lung ultrasound images and finding the best TL technique with U-Net, a convolutional neural network for precise and fast image segmentation. Two approaches of TL were used, using a pre-trained VGG16 model to build the U-Net (V-Unet) and pre-training U-Net network with grayscale natural salient object dataset (X-Unet). Visual results and dice coefficients (DICE) of the models were compared. X-Unet showed more accurate and artifact-free visual performances on the actual mask prediction, despite its lower DICE than V-Unet. A partial-frozen network fine-tuning (FT) technique was also applied to X-Unet to compare results between different FT strategies, which FT all layers slightly outperformed freezing part of the network. The effect of dataset sizes was also evaluated, showing the importance of the combination between TL and data augmentation.