Investigation of Class Separability Within Object Detection Models in Histopathology

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Deep learning for medical imaging

Background:

• Object detection is crucial for histopathological image analysis, but model generalization for clinical use is challenging.
• Existing evaluation metrics lack insight into model failures, especially under domain shifts.
• Understanding model behavior is key to improving deep object detection in histopathology.

Purpose of the Study:

• To propose a novel quantitative method for assessing the discriminative power of a model's latent space in object detection.
• To provide insights into model failure mechanisms and guide architectural optimization.
• To enhance the clinical applicability of deep object detection models in histopathology.

Main Methods:

• Developed a quantitative method applicable to object detection models with local correspondences (e.g., RetinaNet, FCOS, YOLO).
• Adapted and evaluated two metrics: generalized discrimination value and Hellinger distance.
• Incorporated these metrics to trace discrimination across model layers and coordinates.

Main Results:

• Empirically validated the method on real-world histopathology datasets.
• Demonstrated the method's effectiveness in capturing model discrimination properties.
• Showcased its utility in providing insights for architectural optimization.

Conclusions:

• The proposed method offers a novel way to evaluate the discriminative power of latent spaces in object detection models.
• It bridges the gap between performance evaluation and understanding model behavior in histopathology.
• This contributes to more robust and clinically applicable deep learning models for medical image analysis.