An AI-based system for medico-legal document analysis

Medico-legal documents produced in hospital compensation claim procedures are heterogeneous, non-standardized, and often of poor visual quality, making manual analysis slow and error-prone. In this work, we present an end-to-end Document AI pipeline that integrates Document Layout Analysis (DLA) and Optical Character Recognition (OCR) to support the forensic medicine department of Careggi University Hospital in Florence. We construct a real-world dataset of 60 legal medicine documents (809 pages) and adopt the DocLayNet taxonomy to annotate structural components. For DLA, we fine-tune YOLOv8m on the annotated corpus, achieving an mAP@50 of 0.633 and mAP@50-95 of 0.452. For OCR, we compare Tesseract with three Qwen3-VL multimodal models and show that, after lightweight LoRA fine-tuning, Qwen3-VL reduces WER and CER by at least 74.0% and 90.9%, respectively, significantly outperforming the baseline. The resulting pipeline produces structured, machine-readable representations of medico-legal documents and lays the foundation for downstream tasks such as case retrieval, de-identification, and decision support. This study demonstrates the feasibility and benefits of applying modern Document AI and vision-language models to medico-legal workflows in real-world settings.