Research map / Probabilistic and scientific modeling / Molecules and proteins
A Trusted Lesion-assessment Network for Interpretable Diagnosis of Coronary Artery Disease in Coronary CT Angiography
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TLA-Net, a multi-view (global 3D cube plus local 2D slices) CCTA network that uses causal intervention for evidence collection and Dirichlet-based clinically-aligned uncertainty integration to diagnose coronary stenosis and plaque type interpretably.
Problem
Automated CAD lesion assessment from CCTA lacks interpretability and clinically reliable predictions: confounders corrupt the imaging-to-prediction causal path, and diagnoses come without confidence estimates, which hurts generalization across clinical centers.
Approach
Builds a causal graph of the diagnostic process and uses an intervention P(Y|do(X)) computed over a multi-view intervention medium (NWGM approximation), then hierarchically fuses per-view Dirichlet evidence from global-to-local views via Dempster's rule to output calibrated uncertainty.
Key results
- Artery-level ACC of 0.953 on proportional split (DS1) for stenosis degree, versus 0.934 for the best SOTA (Zhang, Ma, and Li 2022).
- Center-based split (DS2) stenosis-degree ACC of 0.913 vs 0.875 for the best prior method (Van Herten et al. 2023).
- Uncertainty integration (CUI) reaches AUC 0.963 and ECE 0.159 on DS1, versus 0.938 AUC and 0.188 ECE for dropout.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use causal intervention P(Y|do(X)) instead of self-attention computation (ablation) | TLA-Net CEC · DS1 and DS2 lesion assessment | performance | Helped |
| integrate multi-view uncertainty using clinical-prior-based CUI rather than Softmax probabilities alone (ablation) | TLA-Net · DS1 and DS2 artery-level | ACC: 0.902 / 0.887 -> 0.953 / 0.944 | Helped |
| integrate multi-view uncertainty using clinical-prior-based CUI rather than Softmax probabilities alone (ablation) | TLA-Net · DS1 and DS2 artery-level | AUC: 0.921 / 0.907 -> 0.963 / 0.950 | Helped |
| integrate multi-view uncertainty using clinical-prior-based CUI rather than Softmax probabilities alone (ablation) | TLA-Net · DS1 and DS2 artery-level | ECE: 0.217 / 0.247 -> 0.159 / 0.173 | Helped |
| use CUI instead of Dropout uncertainty estimation (ablation) | TLA-Net · DS1 and DS2 artery-level | ACC: 0.911 / 0.889 -> 0.953 / 0.944 | Helped |
| use CUI instead of ensemble uncertainty estimation (ablation) | TLA-Net · DS1 and DS2 artery-level | ACC: 0.926 / 0.906 -> 0.953 / 0.944 | Helped |
| use both global and local views rather than only global or only local views (ablation) | TLA-Net · lesion assessment | performance | Helped |
| select 9 local views rather than fewer local views (ablation) | TLA-Net · lesion assessment | performance | Helped |
| retain views centered on the coronary centerline (ablation) | TLA-Net · lesion assessment | performance | Helped |
| use TLA-Net instead of Zhang, Ma, and Li 2022 | artery-level stenosis degree diagnosis · DS1 | ACC: 0.934 -> 0.953 | Helped |
| use TLA-Net instead of Van Herten et al. 2023 | artery-level plaque component diagnosis · DS2 | ACC: 0.851 -> 0.909 | Helped |
Limitations
- Evaluated only on in-house data from two clinical centers with CCTA alone; no multimodal input.
- Relies on coronary centerline extraction and CPR reconstruction, whose inaccuracies the method only partially addresses via random voxel shifts.
Experimental setup
Datasets: In-house CCTA dataset from two clinical centers (CC1: 218 patients, 320-row CT; CC2: 352 patients, dual-source CT)
Benchmarks: In-house CCTA lesion dataset (proportional split DS1) In-house CCTA lesion dataset (center-based split DS2)
Metrics: ACC Precision Recall F1 NPV Specificity AUC ECE
Baselines: Tejero-de Pablos et al. 2019 Zreik et al. 2018 Denzinger et al. 2019 Ma et al. 2021 Zhang, Ma, and Li 2022 Van Herten et al. 2023 Softmax Monte-Carlo dropout Ensemble
Builds on: Dirichlet evidential uncertainty (Sensoy et al. 2018) Shift Transformer (Swin-style 3D) Dempster-Shafer combination rule backdoor/front-door causal intervention
Closest papers by meaning
- TumorChain: Interleaved Multimodal Chain-of-Thought Reasoning for Traceable Clinical Tumor Analysis (ICLR 2026 · similarity 0.84)
- CMR-RD: Long-Tailed Adaptive VLM for Explainable CMR Diagnosis (CVPR 2026 · similarity 0.82)
- CardioLens: Revealing the Clinical Reality Gap of MLLMs via Multi-Sequence Cardiac MRI Evaluations (NeurIPS 2026 · similarity 0.81)
- Rethinking Visual Attribution for Chest X-ray Reasoning in Large Vision Language Models (NeurIPS 2026 · similarity 0.81)
- DDx-TRACE: A Benchmark for Medical Diagnostic Trajectories in VLMs (NeurIPS 2026 · similarity 0.81)
- CXReasonBench: A Benchmark for Evaluating Structured Diagnostic Reasoning in Chest X-rays (NeurIPS 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.