ConvNext Mitosis Identification-You Only Look Once (CNMI-YOLO): Domain Adaptive and Robust Mitosis Identification in...

Area of Science:

• Digital pathology
• Computational pathology
• Machine learning in oncology

Background:

• Accurate mitosis detection is crucial for cancer diagnosis and prognosis in digital pathology.
• Challenges include cell morphology variability and domain shift, hindering model generalization.
• Existing methods struggle with robustness across different data sources.

Purpose of the Study:

• To develop a robust deep learning model for accurate mitosis identification in histopathological images.
• To improve the generalization capability of mitosis detection models across various cancer types, scanners, and species.
• To enhance cancer diagnosis and prognosis through improved mitotic cell detection.

Main Methods:

• Introduced ConvNext Mitosis Identification-You Only Look Once (CNMI-YOLO), a two-stage deep learning approach.
• Utilized YOLOv7 for cell detection and ConvNeXt for cell classification.
• Validated the model on the Mitosis Domain Generalization Challenge 2022 dataset and external test sets.

Main Results:

• CNMI-YOLO achieved a superior F1 score of 0.795 on the Mitosis Domain Generalization Challenge 2022 dataset.
• Demonstrated robust generalization with F1 scores of 0.783 on melanoma and 0.759 on sarcoma test sets.
• Outperformed existing models in precision, recall, and F1 score, with strong performance on unseen data.

Conclusions:

• The CNMI-YOLO model offers a significant advancement in automated mitosis detection for digital pathology.
• The model exhibits strong robustness and generalization capabilities, suitable for real-world clinical applications.
• This approach holds potential for improving the accuracy and efficiency of cancer diagnosis and prognosis.