Explainable Cross-Disease Reasoning for Cardiovascular Risk Assessment from LDCT

TL;DR

This work tackles the problem of opportunistic cardiovascular risk assessment from LDCT by proposing an Explainable Cross-Disease Reasoning Framework that mirrors clinical reasoning. The architecture combines a pulmonary perception module, a knowledge-guided pulmonary-to-cardiac reasoner, and a cardiac region extractor, fused with a multimodal predictor to yield physiologically grounded CVD risk while producing interpretable explanations. It achieves state-of-the-art discrimination on CVD screening and mortality prediction in NLST and demonstrates robust ablations and visualizations (Grad-CAM and textual attribution) that align with known cardiopulmonary mechanisms. The practical impact lies in enabling a single LDCT to deliver dual-purpose, transparent risk stratification, advancing radiology AI toward integrated, mechanism-aware clinical decision support.

Abstract

Low-dose chest computed tomography (LDCT) inherently captures both pulmonary and cardiac structures, offering a unique opportunity for joint assessment of lung and cardiovascular health. However, most existing approaches treat these domains as independent tasks, overlooking their physiological interplay and shared imaging biomarkers. We propose an Explainable Cross-Disease Reasoning Framework that enables interpretable cardiopulmonary risk assessment from a single LDCT scan. The framework introduces an agentic reasoning process that emulates clinical diagnostic thinking-first perceiving pulmonary findings, then reasoning through established medical knowledge, and finally deriving a cardiovascular judgment with explanatory rationale. It integrates three synergistic components: a pulmonary perception module that summarizes lung abnormalities, a knowledge-guided reasoning module that infers their cardiovascular implications, and a cardiac representation module that encodes structural biomarkers. Their outputs are fused to produce a holistic cardiovascular risk prediction that is both accurate and physiologically grounded. Experiments on the NLST cohort demonstrate that the proposed framework achieves state-of-the-art performance for CVD screening and mortality prediction, outperforming single-disease and purely image-based baselines. Beyond quantitative gains, the framework provides human-verifiable reasoning that aligns with cardiological understanding, revealing coherent links between pulmonary abnormalities and cardiac stress mechanisms. Overall, this work establishes a unified and explainable paradigm for cardiovascular analysis from LDCT, bridging the gap between image-based prediction and mechanism-based medical interpretation.

Figures (5)

• Figure 1: Low-Dose Chest CT as a Dual-Purpose Imaging Modality for Cardiopulmonary Assessment. A single LDCT scan inherently captures both pulmonary and cardiac anatomy. (a) Example case showing multiple pulmonary abnormalities including emphysema, fibrosis, and bronchiectasis. (b) Representative case highlighting bronchiectasis in the lower lobes. (c) Case with small lung nodules and surrounding parenchymal changes. (d) Upper-lobe emphysema and fibrotic sequela near the apical region. Each row displays axial, coronal, and sagittal views, demonstrating that LDCT encompasses diverse pulmonary pathologies along with the central cardiac region, motivating our unified framework for joint cardiopulmonary risk prediction.
• Figure 2: Overview of the proposed explainable cross-disease reasoning framework for cardiopulmonary risk assessment from LDCT. A single low-dose chest CT (LDCT) volume serves as the unified input for both pulmonary and cardiovascular evaluation. The framework comprises three coordinated branches: (i) the lung analysis module estimates longitudinal malignancy risk over $T$ years, providing a pulmonary risk trajectory that reflects disease progression; (ii) the pulmonary-to-cardiac reasoner converts structured lung findings into intermediate cardiovascular indicators and natural-language rationales, forming an interpretable bridge between pulmonary abnormalities and cardiac mechanisms; and (iii) the cardiac feature extractor localizes a compact 3D subvolume around the heart to learn morphological and structural biomarkers. The resulting embeddings—pulmonary risk, reasoning-derived indicators, and cardiac representations—are fused through a multimodal prediction head to generate subject-level cardiovascular outcomes for both screening and mortality assessment. This cross-disease integration enables physiologically grounded reasoning from lung pathology to cardiovascular risk, providing both predictive accuracy and transparent interpretability.
• Figure 3: Architecture of the proposed Agentic Pulmonary-to-Cardiac Reasoning Module. Given an LDCT scan, the Pulmonary Perception Agent first summarizes thoracic abnormalities into a structured set of pulmonary findings (bottom left). These findings are then processed by the Knowledge Reasoning Agent, which recalls domain knowledge to associate observed findings with relevant pathophysiological mechanisms (e.g., hypoxemia, systemic inflammation, hemodynamic stress). Next, the CVD Diagnostic Reasoning Agent integrates these inferred mechanisms into higher-level cardiopulmonary effects such as pulmonary hypertension or heart failure risk, generating a concise natural-language explanation (right). The resulting dual output—structured indicators and textual rationale—provides both machine-interpretable representations for downstream inference and human-verifiable reasoning for clinical validation.
• Figure 4: Receiver operating characteristic (ROC) curves for CVD screening (left) and CVD mortality prediction (right) across progressively enhanced variants. "Cardiac Region Only" uses localized 3D cardiac features; "+ Lung-Risk" adds malignancy-derived pulmonary risk; "+ Lung-Risk & Reasoning" further introduces pulmonary-to-cardiac reasoning. Both tasks show consistent gains as modules are added, with the final variant achieving the steepest and most left-shifted curves, indicating superior sensitivity–specificity trade-offs.
• Figure 5: Visualization of explainable cross-disease reasoning. Left: Grad-CAM activation map over the 3D cardiac subvolume highlighting structural regions most indicative of cardiovascular risk. Right: textual attribution heatmap from the pulmonary-to-cardiac reasoning pathway, emphasizing key phrases associated with hemodynamic stress and vascular remodeling. The complementary focus of both modalities illustrates how the framework aligns structural and physiological evidence for interpretable cardiopulmonary prediction.