Automated generation of radiology reports has the potential to improve diagnostic consistency while reducing the burden on clinicians. This study investigates four deep learning approaches for breast ultrasound report generation using the BrEaST dataset, which contains 256 annotated images with lesion masks, BI-RADS descriptors, and histopathology labels. The models compared include a baseline CNN–Transformer, a BERT-enhanced variant, a BioBERT-based model, and a vision-only encoder–decoder. Performance is assessed through BLEU, METEOR, ROUGE-L, and CIDEr, together with Grad-CAM visualizations for interpretability. The BioBERT model produces the strongest results, reaching a ROUGE-L score of 0.6146 and a CIDEr score of 0.4447, while the vision-only system performs worst across all measures (BLEU-4 = 0.2702). Grad-CAM inspection shows that BioBERT aligns more closely with lesion regions, suggesting better clinical grounding. These findings indicate that combining lesion-aware image encoding with biomedical language pretraining enhances both the semantic quality and interpretability of automated breast ultrasound reports.