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

Object Automated report generation (RG) from 3D Time-of-Flight magnetic resonance angiography (TOFMRA) is clinically valuable but technically challenging due to the complex vascular topology. Existing approaches often capture the global context of brain vasculature but fail to incorporate detailed vascular topology, resulting in missed detections of small or morphologically subtle intracranial aneurysms (IAs). To address this, we propose ToRep, a topology-guided RG model designed to enhance both aneurysm detection and report quality.
Methods We retrospectively collected TOF-MRA scans from 1,100 patients diagnosed with IAs between 2021.12 and 2024.04 under IRB approval. From unstructured radiology reports, clinically relevant sentences describing aneurysm, stenosis, and occlusion were extracted. The dataset was divided into training, validation, and test sets (3:1:1). In Fig.1, ToRep employs a dual-branch encoder: one branch encodes the volumetric MRA context, while the other captures vascular topology from precomputed vessel maps. These two representations are fused through a topology-guided cross-attention (TCA) mechanism (Fig.2), which enhances anatomical awareness. The fused features are subsequently passed to a Transformer-based encoder-decoder (CTViT) [1] for RG. Training was conducted using token-level cross-entropy loss restricted to valid tokens only.
Results Performance was evaluated regarding language quality (BLEU-1/2/3/4, ROUGE-L) and clinical accuracy of aneurysm-related findings (accuracy, sensitivity, specificity, precision, recall, F1-score, and intraclass correlation coefficient). Compared with two baselines, CTViT (input: MRA-only) and CTViT+Vessel (inputs: MRA with vessel maps), ToRep consistently achieved higher scores in both fluency and clinical accuracy. Particularly, ToRep generated semantically precise reports (Fig.3) and demonstrated superior localization and size estimation of IAs, which were often misrepresented by baselines (Fig.4&5).
Conclusions We present ToRep, a topology-aware RG framework that effectively incorporates vascular structure to improve the detection and description of IAs in 3D ToF-MRA. Clinically, ToRep shows the potential to reduce radiologists’ diagnostic workload, minimize missed detections of IAs, and support consistent documentation in large-scale screening environments.