Deep learning algorithms out-perform veterinary pathologists in detecting the mitotically most active tumor region.

Area of Science:

• Veterinary Pathology
• Computational Pathology
• Artificial Intelligence in Oncology

Background:

• Manual mitotic figure counting is crucial for tumor grading but suffers from inter-rater variability.
• Uneven mitotic figure distribution within tumor sections significantly impacts count accuracy based on selected areas.

Purpose of the Study:

• To evaluate the impact of area selection on manual mitotic counts in canine cutaneous mast cell tumors.
• To compare the performance of deep learning models against veterinary pathologists in assessing mitotic density.
• To explore computer-based area selection for improving inter-rater agreement in mitotic counting.

Main Methods:

• Eight veterinary pathologists manually selected regions of interest for mitotic counting on 32 whole slide images.
• Three deep learning approaches were evaluated: direct patch prediction, segmentation mask derivation, and a two-stage object detection pipeline.
• Performance was assessed by comparing model predictions to ground truth mitotic counts and pathologist performance.

Main Results:

• Deep learning models, on average, outperformed human experts in mitotic count prediction.
• The two-stage object detection model achieved the highest correlation with ground truth mitotic counts (0.963-0.979).
• Significant variability in area selection was observed among pathologists, contributing to high variance in manual counts.

Conclusions:

• Computer-based area selection can support pathologists, potentially reducing inter-rater variability in manual mitotic counts.
• Deep learning offers a promising avenue for objective and accurate mitotic density assessment in tumor grading.
• Objective area selection via AI could enhance the reliability of mitotic counts in histopathological diagnostics.