Clinical Scorecard: Patient-Centric Evaluation of CNN and Transformer Models for Classifying Breast Cancer Histopathology Images

At a Glance

Key Highlights

• Nine deep learning architectures evaluated on the BreaKHis dataset.
• Strict 5-fold patient-aware cross-validation protocol used.
• ResNet50 achieved the highest mean accuracy (0.9267 ± 0.0435).
• No statistically significant differences in performance among models.
• Intermediate magnification levels provided better feature discrimination.

Guideline-Based Recommendations

Diagnosis

• Utilize automated deep learning models to assist in histopathological diagnosis.

Management

• Implement patient-aware evaluation protocols to avoid data leakage.

Monitoring & Follow-up

• Regularly assess model performance using standardized metrics.

Risks

• Be cautious of over-optimistic performance estimates due to data leakage.

Patient & Prescribing Data

Patients diagnosed with breast cancer requiring histopathological evaluation.

Automated systems can enhance diagnostic efficiency and consistency.

Clinical Best Practices

• Adopt rigorous benchmarking frameworks for model evaluation.
• Ensure uniform training conditions across different model architectures.
• Focus on reproducibility in model performance assessments.

References

• BreaKHis Dataset
• Deep Learning in Medical Image Analysis