Leakage-Free Multimodal Fusion of Radiomics and 3D Deep Learning for CT-Based PDAC Classification

DOI

10.2991/978-94-6239-713-2_4How to use a DOI?

Keywords

Pancreatic ductal adenocarcinoma; computed tomography; radiomics; 3D deep learning; multimodal fusion; nested cross-validation; explainable AI

Abstract

PDAC is inherently lethal, which can be mostly explained by late detection and low sensitivity of visual inspection during contrast-enhanced CT. Minor changes in the textures, isoattenuation lesions, and large inter-observers’ variability still remain obstacles to consistent detection. AI-based methods have shown real promise here, but single-modality approaches, radiomics alone, or deep learning alone, struggle when scanner protocols vary across sites and patient populations don’t cooperate. Here we propose a no-leakage multi-approach framework for CT-based PDAC classification that takes a combination of handcrafted radiomics, 3D deep learning and decision-based fusion. Radiomics features are extracted from anatomically aligned regions of interest and then harmonized across scanners to account for variability introduced by differences in imaging equipment and acquisition settings. In parallel, a 3D early fusion ResNet uses whole-body CT scans and body part masks to get to know the spatial context of the cancer. To avoid data leakage, deep features are extracted using an out-of-fold strategy before being combined with radiomics features. Two fusion approaches are evaluated: average fusion and stacked fusion. The performance of our framework is evaluated using a very strict protocol of nested cross-validation where area under the ROC curve was used as the primary metric. Across all three pipelines i.e. radiomics, deep features, and fusion it was observed that non-linear models consistently outperformed their linear counterparts. The stacked fusion model reached an AUC of 0.955, the volumetric deep learning model hit 0.96. Both discriminate well between PDAC and non-PDAC cases. The results suggest that combining modalities with strict leakage controls produces a more stable classifier than either approach alone, which matters if the goal is something a radiologist can actually trust.

Copyright

© 2026 The Author(s)

Open Access

Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

Download article (PDF)

TY  - CONF
AU  - Anoushka
AU  - Saurabh Singhal
PY  - 2026
DA  - 2026/06/25
TI  - Leakage-Free Multimodal Fusion of Radiomics and 3D Deep Learning for CT-Based PDAC Classification
BT  - Proceedings of the International Conference on Advances in Computing Technology and Artificial Intelligence (COMPUTATIA 2026)
PB  - Atlantis Press
SP  - 51
EP  - 65
SN  - 2589-4919
UR  - https://doi.org/10.2991/978-94-6239-713-2_4
DO  - 10.2991/978-94-6239-713-2_4
ID  - 2026
ER  -