Automatic Radiology Report Generation via Self-Attention Mask Optimization

Automatic radiology report generation is critical to alleviate radiologists' workload. Existing methods suffer from inadequate cross-modal alignment due to suboptimal self-attention mask. This study proposes an improved R2Gen model using Bidirectional Autoregressive (BAR) mask to replace standard Sequence-to-Sequence (S2S) mask. Experiments on the IU X-Ray and MIMIC-CXR datasets show that the BAR mask achieves state-of-the-art performance: it outperforms in BLEU scores (e.g., BLEU-4: 0.188) and clinical efficacy (F1-score: 0.325), demonstrating enhanced capability in generating terminology.