Deep learning (DL) has revolutionized cancer detection accuracy, speed, and accessibility.
Despite its promise, integrating DL into clinical practice presents substantial challenges, including limitations in data quality and standardization, as well as ethical and regulatory concerns, and the need for model interpretability and transparency.
This review highlighted the current applications, opportunities, and challenges of DL in oncology.
This review also emphasized the importance of interdisciplinary collaboration, the integration of next-generation AI techniques, and the adoption of multimodal data approaches to improve diagnostic precision and support personalized cancer treatment.
Additionally, this review proposed some solutions aiming to solve the current challenges:
- To solve dataset size limitations, future work could involve domain-specific augmentations, such as simulating variations in histopathology staining and synthetic tumor growth patterns.
- In cases where discrepancies are inevitable, we propose that institutions develop and adopt federated learning protocols with federated domain adaptation, enabling models to handle slight dataset distribution shifts between institutions.
- To reassure patients and restore clinical trust, we propose introducing confidence thresholds that flag ambiguous cases for additional human review.
- Furthermore, models should be accompanied by detailed "model cards" explaining their training data, limitations, failure modes, and tested demographic performance.