Pathcap: A Contrastive Learning Approach for Automatic Pathology Report Generation in Liver Cancer Diagnosis

Pathological diagnosis serves as the gold standard for liver cancer assessment; however, the process of generating detailed pathology reports is highly labor-intensive. This study introduces PathCap, a novel contrastive learning-based framework for automatic pathology report generation. We construct SYS-PathCap, the largest Chinese dataset with 1,526 liver whole-slide images (WSIs) paired with pathology reports. PathCap employs a dual-stream Transformer architecture to integrate visual features with spatial awareness and textual embeddings. The model achieves state-of-the-art performance in WSI-based report generation, surpassing conventional methods in BLEU, ROUGE, and METEOR scores. Additionally, fine-tuning experiments demonstrate its effectiveness in liver cancer grading and staging classification, highlighting its potential for clinical applications.