Adaptive Student’s T-Loss with Multi-Scale Uncertainty Modeling for Reliable Polyp Detection

Accurate polyp segmentation in colonoscopy images is critical for early colorectal cancer detection, yet remains challenging due to reflections, occlusions, motion artifacts, and significant inter-and intra-polyp appearance variability.Compounding these difficulties, medical image segmentation often suffers from noisy or inconsistent ground-truth annotations, with inter-annotator Dice coefficients ranging from 0.72 to 0.88 in medical imaging tasks [1].These challenges contribute to the 17-28% polyp miss rate during conventional colonoscopy procedures [2,3].We introduce Student's T-Loss, a novel robust loss function specifically designed for polyp segmentation that addresses these challenges through five key innovations.First, Student's T-Loss incorporates a per-image learnable degrees-of-freedom parameter , predicted by a lightweight NuPredictor network to dynamically adjust robustness to outliers.Second, it employs per-pixel precision weights for spatially adaptive error sensitivity, allowing the model to modulate its response to uncertain regions.Third, it implements a multi-scale aggregation scheme that computes and combines loss at multiple spatial resolutions to capture both coarse structural context and fine-grained details.Fourth, we introduce uncertainty-guided active learning that leverages Student's T-Loss's inherent uncertainty estimates to identify the most ambiguous cases for expert re-annotation, reducing annotation burden by 37% while maintaining performance.Fifth, we implement dynamic threshold adaptation where the segmentation threshold varies spatially based on local uncertainty estimates, improving precision-recall balance by 12.3% in challenging boundary regions.Unlike conventional robust loss functions noise, Student's T-Loss specifically addresses the structured noise patterns inherent in colonoscopy data.Our approach integrates these components within a U-Net architecture with ResNet-34 encoder, where all elements, including segmentation parameters, -predictor, and -maps, are jointly optimized via backpropagation.We evaluated Student's T-Loss across five public polyp segmentation benchmarks: EndoScene, CVC-ClinicDB, ETIS-LaribPolypDB, Kvasir, and CVC-ColonDB.Our method achieved state-of-the-art performance across multiple metrics.Notably, Student's T-Loss obtained the lowest Hausdorff distance across all datasets, with an average reduction of 14.6% compared to standard T-Loss and a remarkable 45.96% reduction on EndoScene.It also achieved the lowest false discovery rate on all five datasets, improving over T-Loss by up to 38.7% on EndoScene and 24.5% on Kvasir.Additionally, Student's T-Loss demonstrated superior calibration, achieving expected calibration error as low as 0.44% on EndoScene and outperforming baselines on four of five datasets.A comprehensive ablation study confirmed the contribution of each component, with the full Student's T-Loss framework showing incremental improvements in both global and local uncertainty modeling.Statistical validation through 20 independent runs with paired t-tests confirmed the significance of performance gains (p<0.01 after Bonferroni correction).Despite these substantial improvements, Student's T-Loss maintains clinical deployability with real-time inference at 46.7 FPS (exceeding the 30 FPS clinical threshold) and minimal computational overhead, adding only 0.026M parameters and 0.102 GB VRAM during training.This balance of accuracy and efficiency makes it particularly suitable for real-world clinical settings.Our work establishes a new paradigm for robust polyp segmentation by jointly addressing annotation noise, domain variability, and boundary uncertainty.The consistent performance improvements across diverse datasets demonstrate Student's T-Loss's potential to reduce polyp miss rates and enhance the reliability of computer-aided detection systems in 118-2 colonoscopy.This approach not only advances polyp segmentation but also provides a blueprint for adaptive loss design in other noisy-label medical imaging domains where trustworthy segmentation is critical for clinical decision-making.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.