TriPACT: A Causality-motivated Multimodal Framework with Active Counterfactual Testing for Radiology Report Generation
Authors: Cheddi, F., Habbani, A., Nait-Charif, H.
Journal: IEEE Access
Publication Date: 01/01/2026
eISSN: 2169-3536
DOI: 10.1109/ACCESS.2026.3701619
Abstract:The increasing volume of medical imaging examinations imposes the need for intelligent systems capable of generating clinically reliable radiology reports, which can support diagnostic workflows and improve reporting efficiency. Despite rapid progress in multimodal learning, current radiology report generation models remain limited by three major limitations, including the lack of structured causally informed reasoning, suboptimal integration of heterogeneous multimodal evidence, and the frequent emergence of clinically unsupported or hallucinated findings. To address these challenges, we propose TriPACT, a causality-motivated multimodal framework that unifies visual, clinical, and textual information through patient-specific structured clinical modeling and counterfactual validation. TriPACT integrates a multi-view visual backbone to extract semantically rich representations from paired frontal and lateral chest X-rays, a dynamic causal clinical graph for encoding structured clinical semantics through clinically grounded and causally motivated relations. A cross-modal fusion and intervention module enforces multi-modal evidence consistency via single-pass counterfactual testing to validate or refute diagnostic inferences based on multimodal evidential support. Finally, a Structured Report Decoder generates section-aware Findings and Impression reports grounded in counterfactually validated evidence. Experiments on public chest X-ray and COV-CTR benchmarks show that TriPACT leads to a significant improvement in factual correctness, a reduction in unsupported statements, and increased robustness under different imaging conditions, advancing the robustness of automatic report generation.
Source: Scopus
TriPACT: A Causality-Motivated Multimodal Framework With Active Counterfactual Testing for Radiology Report Generation
Authors: Cheddi, F., Habbani, A., Nait-Charif, H.
Journal: IEEE ACCESS
Publication Date: 2026
Volume: 14
Pages: 88267-88281
ISSN: 2169-3536
DOI: 10.1109/ACCESS.2026.3701619
Source: Web of Science