Machine learning techniques for mitoses classification.

Area of Science:

• Digital pathology
• Computational biology
• Oncology

Background:

• Accurate cancer staging relies on precise analysis of biopsy material at cellular and structural levels.
• Mitotic figures, as biomarkers of cellular proliferation, are crucial for prognostic information and clinical care.
• Convolutional Neural Networks (CNNs) demonstrate significant potential for improving the accuracy of mitosis detection.

Purpose of the Study:

• To evaluate the performance of state-of-the-art CNN models for mitosis classification in whole slide biopsy images.
• To compare the quantitative performance of ESPNet and DenseNet on a primary melanoma dataset.
• To benchmark four CNN architectures (ESPNet, DenseNet, ResNet, ShuffleNet) on the public MITOS breast cancer dataset.

Main Methods:

• Two CNN models, ESPNet and DenseNet, were trained and evaluated on whole slide skin biopsy images.
• Performance metrics including sensitivity, specificity, and F-score were used for comparison.
• Four CNN architectures were further assessed on the MITOS breast cancer dataset using precision, recall, and F-score.

Main Results:

• On the melanoma dataset, ESPNet and DenseNet achieved high sensitivity (0.976, 0.968) and specificity (0.987, 0.995).
• On the MITOS dataset, DenseNet demonstrated superior F-score (0.927) and recall (0.916) compared to other models.
• ResNet and ESPNet offered faster inference times (6s and 8s, respectively) compared to DenseNet (31s), presenting a trade-off between speed and accuracy.

Conclusions:

• The study successfully evaluated CNNs for mitosis detection in whole slide biopsy images.
• The proposed methodology and architectures are robust and applicable to various biopsy types for mitosis identification.
• CNNs, particularly DenseNet, offer a powerful tool to aid pathologists in improving cancer diagnosis and staging accuracy.