Self-supervised feature learning for invasive ductal carcinoma detection in histopathological images

Invasive Ductal Carcinoma (IDC) is recognized as the most commonly diagnosed type of breast cancer, and its effective distinction based on histopathological images is still a difficult task, as the inter-class variability is high, and there is limited annotated data. To overcome these challenges, the given paper suggests a self-supervised learning-based deep learning architecture to identify IDC at a patch level in histopathological images. The first stage of the suggested method presents contrastive self-supervised pretraining in order to obtain robust and discriminative feature representations on unlabeled histopathological patches through powerful data augmentations. The encoder that has been partially trained is then trained using a supervised method through labeled data to achieve binary classification of the histopathological patches. The proposed model demonstrates exceptional performance, achieving an accuracy of 88.81 %, a precision of 79.39 %, a recall of 1.85, an F1-score of 80.60 %, and an AUC of 0.9511 which shows a successful ability to discriminate between benign and malignant tissue patterns. This framework is assessed using a popular IDC histopathology dataset (IDC regular ps50 idx 5) on patch-wise as well as patient-wise evaluation schemes. The experimental findings prove that the suggested approach can reach competitive classification results and minimize the use of large amounts of labeled data. Moreover, the visualization of attention features is used to highlight informative areas of the histopathological patches aims to provide qualitative insights into the decisionmaking process of the model. Overall, the results indicate that self-supervised pretraining is a scalable and effective approach to representation learning in breast cancer histopathology analysis.