Topology-Aware Report Generation for Intracranial Aneurysm Diagnosis in 3D TOF-MRA

Introduction: Automated report generation (RG) from 3D Time-of-Flight magnetic resonance angiography (TOFMRA) is a clinically valuable but technically challenging task, particularly for diagnosing intracranial aneurysms (IAs). While recent approaches have shown promise in 3D brain imaging tasks, they often overlook critical vascular topologies and struggle to localize subtle aneurysmal changes within the complex cerebral vasculature. To address this, we propose ToRep, a topology-guided RG model incorporating vessel-aware spatial reasoning and structural epresentation. By fusing volumetric appearance and vascular structure through topology-guided cross-attention (TCA), ToRep improves the detection of morphologically subtle IAs often missed by existing approaches while enhancing the generated report's quality.

Methods: We retrospectively reviewed TOF-MRA scans from 1,100 patients diagnosed with IAs between December 2021 and April 2024 under IRB approval. From unstructured radiology reports, clinically relevant sections detailing aneurysm, stenosis, and occlusion were extracted. The dataset was split into training, validation, and test sets (3:1:1). ToRep integrates a dual-branch encoder: one branch encodes the volumetric context of the MRA, while the other captures vessel topology from precomputed vessel maps via a dedicated vessel encoder. To enhance anatomical awareness, we introduce TCA, allowing mutual interaction between contextual and topological features. The topology-fused representation is then passed to a transformer-based encoder-decoder (CTViT) for RG. For training, we employ a token-level crossentropy loss with masking to ensure that the loss computation is restricted to valid (non-padding) tokens only.

Results: Evaluation was performed on both RG quality using BLEU-1/2/3/4 and ROUGE-L, and the accuracy of aneurysm-related findings within generated reports through accuracy, sensitivity, specificity, precision, recall, F1-score, and intraclass correlation coefficient. ToRep was quantitatively and qualitatively compared with two baselines: CTViT (input: MRA-only) and CTViT+Vessel (inputs: MRA and a vessel map). Across all metrics, ToRep consistently outperformed the baselines in both language quality and clinical accuracy. Whereas baselines often failed to produce the report or provided incorrect size and location descriptions, ToRep generated more fluent and semantically accurate reports while also providing precise aneurysm size and localization. These results highlight ToRep’s capability to generate clinically reliable reports by effectively incorporating vessel-aware topological cues.

Discussion & Conclusion: We proposed ToRep to effectively incorporate vascular topology for improved RG related to IAs, thereby enhancing the detection and description of small or morphologically subtle IAs. From a clinical perspective, automatically generated preliminary reports by ToRep could offer meaningful support for radiologists to reduce diagnostic workload, minimize human oversight in detecting subtle aneurysms, and support consistent documentation across large-scale screening settings.