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

Area of Science:

• Histopathology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Accurate mitotic cell identification is vital for cancer grading and prognosis.
• Manual identification of mitotic cells is labor-intensive and time-consuming for pathologists.
• Current automated methods often require extensive labeled data and computational resources.

Purpose of the Study:

• To evaluate the efficacy of foundation models, specifically those trained with Self-Distillation with No Labels (DINO), for mitotic cell detection in histopathology.
• To assess if DINO-based foundation models can achieve performance comparable to traditional supervised models.
• To investigate the resource efficiency of DINO-based models compared to supervised approaches.

Main Methods:

• Utilized foundation models including Vision Transformer (ViT), Cross-Covariance Image Transformers (XCiT), and ResNet-50.
• Employed the Self-Distillation with No Labels (DINO) self-supervised learning method for model pre-training.
• Evaluated model performance on the MIDOG2021, MIDOG2022, and ICPR2014 datasets.

Main Results:

• DINO-based foundation models achieved F1-scores of 0.8174 (MIDOG2021), 0.8275 (MIDOG2022), and 0.7509 (ICPR2014).
• Performance of DINO models was comparable to supervised models, which yielded F1-scores of 0.8254, 0.8390, and 0.7884 on the respective datasets.
• Foundation models demonstrated significant resource efficiency compared to supervised methods.

Conclusions:

• Foundation models trained with DINO offer a viable and efficient alternative for automated mitotic cell detection in histopathology.
• The self-supervised approach reduces the need for extensive labeled data and costly training phases.
• This study highlights the potential of foundation models to improve workflow efficiency in computational pathology.