From Global Radiomics to Parametric Maps: A Unified Workflow Fusing Radiomics and Deep Learning for PDAC Detection

A research team led by Zengtian Deng has introduced a novel AI framework that unifies radiomics and deep learning for improved pancreatic ductal adenocarcinoma (PDAC) detection. Published on arXiv (2602.17986), their work addresses a key limitation in existing fusion approaches that typically leverage only global radiomic features, overlooking the complementary value of spatially resolved radiomic parametric maps.

The technical innovation lies in a two-stage process: first selecting discriminative radiomic features, then injecting them into a radiomics-enhanced nnUNet architecture at both global and voxel levels. This dual-level integration allows the model to leverage both holistic statistical features and localized spatial patterns from medical images. On the benchmark PANORAMA dataset, the framework achieved impressive metrics with AUC=0.96 and average precision (AP)=0.84 in cross-validation testing. Even more compelling was its performance on an external in-house cohort, where it maintained strong results with AUC=0.95 and AP=0.78, outperforming the baseline nnUNet model.

The research demonstrates that handcrafted radiomics, when properly integrated at multiple levels, provide complementary signals to deep learning models for complex medical imaging tasks. The framework's second-place ranking in the competitive PANORAMA Grand Challenge validates its practical effectiveness. This work represents a significant step toward more robust AI-assisted diagnosis tools that combine the interpretability of traditional radiomics with the power of modern deep learning architectures.