Deep Learning-Based Cancer Detection in Histopathological Images of Breast Pathology

Abstract

On a global scale there are an estimated 2.3 million new cases of breast cancer per year. For improvement of patient outcomes and increase of the survival rate, it is essential to detect early and diagnose accurately. Although considered the gold standard for diagnosis, histopathological examination in itself is both time-consuming and is subject to inter observer variability, which may affect a poor consistency and accuracy when terminated by humans and this constitutes a problem. This research flls these well known gaps by developing a systematic deep learning framework to automate the classification and segmentation of breast cancer histopathological images. The framework consists of two main components in the form of tools to help pathologists in their diagnostic workflow. The frst approach leveraged a standalone Vision Transformer (ViT) architecture to process image patches as sequences in order to realize global dependencies. The second approach used a hybrid model of the VGG16 convolutional neural network augmented with a Data efcient Image Transformer (DeiT). This hybrid architecture leverages the local features captured by convolutions and global context modeling abilities of transformers. The hybrid model outperformed the ViT-only model with an accuracy of 95% compared to 91%. This supports the fact that architectural integration is worth pursuing for improving diagnostic precision. Specifically, the Trans U-Net architecture was used in the segmentation part of the framework to identify and delineate histopathology-specific structures. The model was trained to differentiate among Invasive carcinoma, In Situ carcinoma and Benign tissue. To train and evaluate the segmentation model, a set of pixel-wise annotated histopathological images was used. Excellent results were demonstrated with a mean Dice similarity coefficient value of 92%, showing good overlap between predicted and ground truth segmentations. More metrics are marked with high values of Intersection over Union (IoU), sensitivity and specificity. This framework may provide pathologists with a powerful tool for clinical implementation as a decision support system to improve diagnostic efciency, consistency and accuracy. Future work will consist of integration into existing laboratory information systems to maximize clinical impact and adoption.