Uncertainty-aware polyp segmentation with Monte Carlo dropout for trustworthy cross-domain colorectal cancer detection

Abstract

The detection of polyps during colonoscopy plays the vital role in colon cancer prevention, but the majority of deep learning segmentation algorithms do not provide interpretable confidence estimates and significant evaluation across clinical populations. We suggest an uncertainty-aware polyp segmentation model that will use EfficientNet-B4 encoder and UNet++ decoder, where weighted Dice and Focal loss should be used. Monte Carlo Dropout inference hallucinates pixel uncertainty maps, which can be used to risk-stratify clinical visualization. It was zero-shot tested and trained on Kvasir-SEG without fine-tuning and on CVC-ClinicDB. The framework was similar to the Dice with similarity coefficient of 0.864 ± 0.167 on internal and 0.713 ± 0.330 on external test, indicating 17.5 domain generalization gap. The error-uncertainty correlation was statistically significant (Pearson r = 0.335, p < 0.001) with reviewing 20 per cent of high-uncertainty regions leading to 60 per cent of the segmentation errors. Deterministic inference attained 41.97 frames per second, which is suitable to clinical deployment in real-time. Combining the approaches of measuring uncertainty and cross domain assessment results in a clinically feasible segmentation framework that achieves an accurate, interpretable and efficient result. Although further refinements of calibration are still required, the suggested methodology brings credible medical artificial intelligence to closer to practice