Related Experiment Videos
Unified Multi-task Learning for Colorectal Cancer Diagnosis via Uncertainty-aware Routing, Cross-task Consistency,
Divya Midhunchakkaravarthy1, G Muni Nagamani2, Lakshman Narayana V3
1Department of Electronics and Communication Engineering, Lincoln University College, Petaling Jaya, 47301, Malaysia.
Introduction:
The study aims to develop a unified multi-task learning framework for colorectal cancer diagnosis using whole-slide histopathology images. Specifically, it targets joint tissue segmentation and tumor grading, enhancing label efficiency, robustness, and generalization, while minimizing the redundancy and annotation burden typically associated with treating these tasks independently.
Methods:
1. Hierarchical Uncertainty-Gated Task Routing (HUGTR): Dynamically allocates encoder features to decoders based on aleatoric and epistemic uncertainty. 2. Cross-Task Consistency Attention Matrix (CTCAM): Enforces spatial coherence between segmentation and grading by aligning attention maps across tasks. 3. Adaptive Label Denoising with Structural Priors (ALDSP): Employs graph convolutional autoencoders guided by structural priors from tissue segmentation. 4. Contrastive Segmentation-Grading Latent Embedding (CSGLE): Implements a two-headed contrastive learning module to align the latent representation spaces of segmentation. 5. Curriculum-Based Multi-Resolution Task Cascade (CMRTC): Trains the model progressively from low to high resolution images, adapting it to different spatial scales and complexity levels in whole-slide images (WSIs).
Results:
6.8% increase in tumor grading AUC, +3.5% improvement in segmentation Dice score, and 27% reduction in model parameters,14.3% decrease in inter-observer variability Discussion: The combination of uncertainty-aware routing, cross-task alignment, and label denoising significantly enhances both diagnostic precision and model efficiency. By treating segmentation and grading as interrelated rather than isolated tasks, the model better captures shared pathological patterns and domain priors. The incorporation of contrastive learning and multi-resolution training further supports generalization across patients and datasets.
Conclusion:
This unified multi-task framework sets a new benchmark in histopathological analysis for colorectal cancer by effectively integrating tissue segmentation and tumor grading. The method's innovations enable better use of annotations, improved diagnostic consistency, and enhanced scalability, positioning it as a robust AI tool in pathology workflows.
