Polyp-D2ATL beats SOTA in colorectal polyp classification with 82% accuracy

The paper presents Polyp-D2ATL (Deep Domain-Adaptive Transfer Learning), a novel framework designed to improve colorectal polyp classification under challenging real-world conditions such as imbalanced data, label distribution shift, and cross-modality differences. Built on the NICE classification system, Polyp-D2ATL incorporates a custom training strategy that effectively adapts knowledge from source domains while mitigating distribution mismatches. Extensive experiments on the PICCOLO validation and test sets demonstrate consistent improvements across multiple metrics: 82.38% accuracy, 77.49% Macro-F1, and 87.47% specificity on the validation set, with similar gains on held-out test data. The results confirm the model's generalization capacity and potential for clinical deployment.

The work addresses a critical gap in automated polyp diagnosis—many existing models fail when faced with difficult-to-predict polyps or when applied to new imaging modalities. By employing domain-adaptive transfer learning, Polyp-D2ATL maintains robust performance despite shifts in label distributions between training and deployment environments. The authors emphasize the framework's ability to handle both balanced and imbalanced datasets, a common pain point in medical AI. With 15 pages of detailed methodology and five figures illustrating the training pipeline and results, the paper offers a comprehensive solution for improving early and accurate colorectal cancer screening. The arXiv submission (2606.15000) provides full access to the code and data for replication.