Machine learning in oral squamous cell carcinoma: Current status, clinical concerns and prospects for future-A systematic review.

Integrating deep convolutional neural networks with marker-controlled watershed for overlapping nuclei segmentation in histopathology images

Automated segmentation of cell nuclei is crucial for the early diagnosis of cancer as the characteristics of the cell nuclei are mainly associated with the assessment of malignancy. Only a few research work has been done on automated segmentation of cell nuclei on cytology pleural effusion images, which is poorly handled by previous methods. In addition, cytology pleural effusion image itself is still challenging due to the poor contrast of images, a variety of cells, and overlapping cells. To deal with the remained problems, this paper presented the algorithm for automated segmentation of cell nuclei in pleural effusion cytology images which contain touching and overlapping cells. First, the preprocessing step is carried out to reduce noise and enhance the contrast using the median filter and CLAHE respectively. The cell nuclei are segmented using K Mean Clustering algorithm in LAB color space. Then, the boundaries are corrected and non-nuclei regions are eliminated by the morphological operations. Finally, the overlapping cell nuclei are isolated depending on the watershed method, subsequently, boundaries of isolated cell nuclei are estimated using ellipse fitting method. The proposed system is evaluated on the local dataset containing 35 images of cytology pleural effusions with normal benign and cancer cells. The experimental results yield the accuracy of Precision= 0.90, Recall=0.89, F-meas-ure=0.89, Dice Similarity Coefficient=94% and Piccard In-dex=89% respectively. The obtained results are verified and compared with the ground truth images manually annotated by experts.

K mean clustering based automated segmentation of overlapping cell nuclei in pleural effusion cytology images

Stratified squamous epithelial biopsy image classifier using machine learning and neighborhood feature selection

Multi-Objective Evolutionary Algorithms (MOEAs) are known to solve problems where two or more conflicting goals are involved. To accomplish it, MOEAs incorporate strategies to determinate optimal trade-offs between each of the objective functions. In this paper, an Unassisted image Thresholding (UTH) methodology is proposed based on MOEAs. UTH takes advantage of the trade-off mechanisms present on MOEAs to perform the image thresholding while simultaneously determinating the number thresholds required to segment each image and the best placement of each threshold along the histogram of the image. The image thresholding problem is commonly addressed as the search for the best possible thresholds able to partition a given image into a finite number of homogeneous classes. Such approach requires the assistance of a designer to determinate the number of threshold values that will properly segment the image. However, as images can vary significantly, the definition of an optimal number of thresholds should be performed for each image. Thus, a methodology able to determinate both the number of thresholds and the best placement of each value contributes to a general segmentation scheme. In the proposed approach, UTH redefines the thresholding problem as a multi-objective task with two conflicting goals. The first goal is the quality of the segmented image, and it is computed as a non-parametric criteria to evaluate candidate threshold points. The second goal is the normalized number of threshold points. Since the number of thresholds is not fixed, a particle encoding the thresholds with variable length is used. The strategy of UTH is coupled with three MOEAs namely NSGA-III, PESA-II and MOPSO using as the non-parametric criteria the Cross Entropy. According to the results, the UTH NSGA-III formulation outperforms UTH-PESA-II and UTH-MOPSO regarding convergence and quality of the resulting image.

Unassisted thresholding based on multi-objective evolutionary algorithms

In this paper, image segmentation process using thresholding technique is discussed We focus our study for nuclei detection of Hematoxilin Eosin (HE) stained colon tissue for the cancer detection Detection of cancer cell at early stage is very important and several studies are being carried out In medical imaging for the processing of microscopic tissue images and especially the detection of cell nuclei is done more and more using digital imagery techniques However, there are many other techniques varies in complexity and quality of detection In this paper the result of different thresholding techniques are applied on HE-stained colon tissue to detect cell nuclei in the image In the proposed method, a binary mask is formed which is of the same size as that of original image In this binary image, the centre of cell nuceli is represented by white colour and the background in black This image is superimposed over the original image The outcome of this is a segmented image to detect the cell nuclei The process is relatively simple and results are encouraging

Image segmentation using thresholding for cell nuclei detection of colon tissue

GPU accelerated stratified squamous epithelium biopsy image segmentation for OSCC detector and classifier

Digital image segmentation is the first step in computer aided diagnostic procedures which are carried out with the help of medical images in the medical field. In this paper, segmentation of Hematoxylin and Eosin (H&E)-stained microscopic image of stratified squamous epithelial layer of oral cavity to separate the squamous cells from the background is performed by different approaches. Due to the complex structure of the squamous epithelial layer, the widely used K-means clustering and thresholding techniques are either not suitable for segmenting such images or unable to furnishthe suitable result. In this work, we are proposing new method for segmentation using Gabor filter. The input image is filtered through a bank of Gabor filters. The number of scales used in constructing the bank of filters is adaptive and automatically computed based on the size of the image. Filtered outputs are taken as 2-dimentional feature vectors. Furthermore, principal component analysis is performed to reduce the dimensionality. In addition, the first principal component is used as the feature image for further processing towards segmentation. This feature image is given as input to both the K-means clustering and thresholding for the final segmentation. The outputs of different approaches are compared. It is found that Gabor filter with thresholding and K-means clustering offers improved result as compared to the conventional ones.

Performance Analysis of Image Segmentation for Oral Tissue

Oral squamous cell carcinoma (OSCC) remains a major death causing oral cancer in developing countries. In recent years, tremendous development in medical imaging devices made microscopic colour images of biopsy samples available to the researchers. Image processing and machine learning techniques can be used to develop automatic cancer grading mechanism. In this work, automatic OSCC classifier using Linear Discriminant Analysis combined with Random Subspace is developed and analyzed. The proposed classifiers automatically classifies the input image in one of the four categories, namely: Normal, Grade-I, II or III. Total 83 colour and texture features are computed from the 100 Haemotoxylin and Eosin (H&E) stained images of oral mucosa. The overall accuracy of the proposed classifier is 93.5% with sensitivity and specificity of 0.89 and 0.95 respectively.

Lakshmi Yamujala

Papers

Stratified squamous epithelial biopsy image classifier using machine learning and neighborhood feature selection

Image segmentation using thresholding for cell nuclei detection of colon tissue

GPU accelerated stratified squamous epithelium biopsy image segmentation for OSCC detector and classifier

Performance Analysis of Image Segmentation for Oral Tissue

Random Subspace Combined LDA Based Machine Learning Model for OSCC Classifier