TACMRN: Triple-Attention Contrastive Memory Retrieval Network for Brain CT Report Generation

Brain CT report generation aims to automatically produce diagnostic reports from 3D brain volumes, reducing radiologists' workload and enhancing clinical decision-making. Despite recent advances, existing methods often struggle to effectively capture comprehensive spatiotemporal features from 3D brain CT data and fail to leverage crucial medical knowledge accumulated from prior cases. To address these challenges, we propose a novel triple-attention contrastive memory retrieval network (TACMRN) that mimics the cognitive workflow of experienced radiologists. Specifically, triple-attention enhancement (TAE) module decomposes the 3D attention into spatial position, spatial channel, and temporal channel branches to efficiently extract fine-grained spatiotemporal features. Multiscale contrastive fusion (MCF) module captures differential features between abnormal and normal images across multiple scales, enhancing lesion perception. Moreover, memory retrieval enhancement (MRE) module introduces learnable memory slots to store and retrieve general medical knowledge, simulating radiologists' reliance on prior experience. Extensive experiments were conducted on the public CTRG-Brain dataset. On CTRG-Brain, the proposed method achieved 0.597 in BLEU-1, 0.503 in BLEU-2, 0.441 in BLEU-3, 0.391 in BLEU-4, 0.291 in METEOR, and 0.540 in ROUGE-L, outperforming all competing methods and achieving state-of-the-art performance, and ablation study validated the effectiveness of each proposed component. The proposed TACMRN provides an effective solution for automated brain CT report generation. By jointly leveraging TAE, MCF and MRE modules, the method substantially captures fine-grained lesion features and improves diagnostic accuracy and reporting consistency. This work offers practical value for the automated radiology report generation, potentially reducing radiologists' workload and accelerating the doctor's diagnosis process.