Automated classification of mitotic catastrophe by use of the centromere fragmentation morphology.

Mitotic catastrophe is a common mode of tumor cell death. Cancer cells with a defective cell-cycle checkpoint often enter mitosis with damaged or under replicated chromosomes following genotoxic treatment. Premature condensation of the under-replicated (or damaged) chromosomes results in double-stranded DNA breaks at the centromere (centromere fragmentation). Centromere fragmentation is a morphological marker of mitotic catastrophe and is distinguished by the clustering of centromeres away from the chromosomes. We present an automated 2-step system for segmentation of cells exhibiting centromere fragmentation. The first step segments individual cells from clumps. We added two new terms, weighted local repelling term (WLRt) and weighted gradient term (WGt), in the energy functional of the traditional Chan-Vese based level set method. WLRt was used to generate a repelling force when contours of adjacent cells merged and then penalized the overlap. WGt enhances gradients between overlapping cells. The second step consists of a new algorithm, SBaN (shape-based analysis of each nucleus), which extracts features like circularity, major-axis length, minor-axis length, area, and eccentricity from each chromosome to identify cells with centromere fragmentation. The performance of SBaN algorithm for centromere fragmentation detection was statistically evaluated and the results were robust.

Patch-based system for Classification of Breast Histology images using deep learning.

In this work, we proposed a patch-based classifier (PBC) using Convolutional neural network (CNN) for automatic classification of histopathological breast images. Presence of limited images necessitated extraction of patches and augmentation to boost the number of training samples. Thus patches of suitable sizes carrying crucial diagnostic information were extracted from the original images. The proposed classification system works in two different modes: one patch in one decision (OPOD) and all patches in one decision (APOD). The proposed PBC first predicts the class label of each patch by OPOD mode. If that class label is the same for all the extracted patches and that is the class label of that image, then the output is considered as correct classification. In another mode that is APOD, the class label of each extracted patch is extracted as done in OPOD and a majority voting scheme takes the final decision about class label of the image. We have used ICIAR 2018 breast histology image dataset for this work which comprises of 4 different classes namely normal, benign, in situ and invasive carcinoma. Experimental results show that our proposed OPOD mode achieved a patch-wise classification accuracy of 77.4% for 4 and 84.7% for 2 histopathological classes respectively on the test set obtained by splitting the training dataset. Also, our proposed APOD technique achieved image-wise classification accuracy of 90% for 4-class and 92.5% for 2-class classification respectively on the split test set. Further, we have achieved accuracy of 87% on the hidden test dataset of ICIAR-2018.