An Effective Multi-Patch Deep Learning Method for Mitosis Detection in Breast Histopathology

Within the realm of computational pathology, the detection of mitotic cells poses a formidable challenge. Many existing approaches rely on hand-crafted features, which often result in poor generalization, as their performance degrades across different tissue types, staining processes, and the various scanners used for digitizing whole slide images. The Multi-Patch Mitosis-Detect (MPMD) framework is suggested in the proposed work to detect mitosis from histology images. To identify mitotic reference regions, the proposed MPMD framework uses the detection module for segmentation by utilizing the Recurrent Residual Convolutional Unit (RRCU). The classification model then employs Inception Recurrence Residual Convolutional Neural Networks (IRR-CNN) to validate the mitotic regions. Furthermore, a novel confidence analysis and the MPMD technique are combined to improve the performance of the detection in the testing phase. The novelty of the proposed multi-patch approach is that: (a) Mean Squared Error (MSE) loss is used instead of Dice Coefficient (DC) loss for both training and testing; (b) Global Average Pooling is used in place of fully connected layers in the classification model to reduce the number of network parameters. Experimental findings demonstrate the performance improvement of the proposed approach compared to existing state-of-the-art methodologies.