Feasibility of Automated Liver Metastasis Detection and Report Generation in Gadoxeotic-enhanced Abdominal MRI using Vision-Language Models

Motivation: To enhance accuracy and efficiency in liver metastasis evaluation in liver MRI. Goal(s): To investigate the feasibility of applying vision-language AI models-(VLM) to hepatobiliary-phase (HBP) MRI for lesion detection, segmentation, and structured report generation. Approach: we refine the vision component of VLM to segment liver and liver lesions on HBP MRI and then prompt the VLM to generate structured reports on these lesions, including their volumetric measurements. Results: The initial VLM shows promising results in lesion detection with patient-level sensitivity of 0.98, and specificity of 0.57. Lesion-segmentation Dice score is 0.78. The VLM is 78% accurate in describing lesion anatomical location. Impact: VLMs offer a promising approach to enhancing both the accuracy and efficiency of liver metastasis evaluation in abdominal MRI interpretation.