Adaptive cross-sequence attention with patient geometry encoding for precise MRI lumbar disease diagnosis

Magnetic resonance imaging (MRI) plays a critical role in identifying lumbar diseases that underlie low back pain (LBP). However, its clinical interpretation remains highly dependent on expert knowledge. Although deep learning has advanced automatic MRI data analysis across various clinical tasks, the accurate diagnosis of lumbar diseases in real clinical applications remains challenging. Compared with other anatomical regions, the lumbar spine exhibits greater structural complexity and often presents subtle pathological manifestations. Meanwhile, clinical MRI data often exhibit pronounced sequence-level heterogeneity across patients due to variations in diagnostic scenarios. These two factors jointly hinder the effective deployment of automatic lumbar disease diagnosis from the pathological and data aspects, respectively. To address these challenges, we propose an adaptive cross-sequence attention framework with patient geometry encoding to accurately identify the underlying causes of LBP. Specifically, critical pathological information is adaptively selected from variable-length MRI sequences through graph-based diagnostic contribution estimation. Beyond image content, geometry prompts derived from patient-centric metadata are employed to compensate for spatial information loss, thereby further facilitating anatomically grounded contextual understanding. To assess the effectiveness of the proposed method, we collected and annotated real-world MRI data from patients with low back pain, establishing a clinically curated multi-center dataset covering seven common chronic lumbar diseases. Extensive experimental results demonstrate that the proposed method consistently outperforms existing approaches across multiple MRI input configurations.