Automatic Generation of Brain Tumor Diagnostic Reports from Multimodality MRI Using Large Language Models

Early detection and precise characterization of brain tumors are critical for improving patient outcomes and survival rates. Magnetic Resonance Imaging (MRI) remains a cornerstone modality in brain tumor diagnostics, offering high-resolution, multi-sequence imaging essential for capturing the complex structure and pathology of intracranial abnormalities. However, the process of diagnostic report generation is labor-intensive, requiring meticulous review, synthesis, and verification by clinicians. In this study, we present a clinical data pipeline, named iReportMed, designed for automated diagnostic report generation specific to brain tumors. The framework consists of a two-phase model. In the initial phase, multi-modality MRI scans are utilized for comprehensive feature extraction and classification, leading to subject-specific labeling. The second phase employ Large Language Models (LLMs) to produce structured diagnostic reports based on extracted features and identified labels. The proposed framework is initially trained on the UCSF-PDGM dataset and subsequently fine-tuned with real clinical data. Results demonstrate that automated diagnostic reporting is both feasible and potentially advantageous for integration into routine clinical work-flows, offering a promising tool for enhancing efficiency in healthcare settings.