Evaluation of Self-Supervised Representation Learning for Mitosis Detection in Histopathological Images

In histopathology, accurate identification of mitotic cells is crucial for classifying specific types of cancer. Pathologists assess tumor aggressiveness by meticulously identifying mitotic cells within defined tissue sections. However, this process is inherently laborious and time-consuming for practitioners. To address this challenge, classical image processing techniques and deep learning-based algorithms have been employed. This study examines the applicability of foundation models developed using the self-supervised learning method, Self-Distillation with No Labels (DINO), for histopathology images, thereby eliminating the need for expensive training phases. Vision Transformer (ViT), Cross-Covariance Image Transformers (XCiT), and Residual Networks with 50 layers (ResNet-50) have been employed as base models. To evaluate the proposed technique, the MIDOG2021, MIDOG2022, and ICPR2014 datasets have been utilized, yielding F1-scores of 0.8254, 0.8390, and 0.7884 for the supervised model and 0.8174, 0.8275, and 0.7509, for DINO model, respectively. An analysis of these results reveals that foundation models demonstrate a performance comparable to supervised models. Since foundation models require fewer computational resources than supervised models, this study also contributes to resource efficiency.