Impact of Annotation Level on Multisequence MRI Models for Preoperative Microvascular Invasion Prediction in Hepatocellular Carcinoma.

Purpose To evaluate the performance of deep learning models integrating multimodal data for predicting microvascular invasion (MVI) in hepatocellular carcinoma and to investigate the impact of different manual annotation methods on performance. Materials and Methods Patients with hepatocellular carcinoma from three institutions were included in this retrospective study; postoperative histopathology served as the reference standard for MVI. Patients from center A were divided into training and internal test sets; patients from centers B and C formed the external test set. Two manual annotations (voxel-level masks, bounding boxes) were performed on MRI scans. Deep learning models were developed using multimodal data. Model performance was evaluated using the receiver operating characteristic, calibration, and decision curve analysis, with area under the receiver operating characteristic curve (AUC) differences tested by the DeLong test. Results A total of 281 patients were included in this study (mean age, 59.05 years ± 11.92 [SD]; 238 male). Single-sequence models achieved internal test AUCs of 0.57-0.76. Multisequence models reached AUCs of 0.86 (95% CI: 0.77, 0.95) with masks and 0.83 (95% CI: 0.73, 0.94) with bounding boxes. Multimodal fusion improved performance (mask: AUC, 0.88 [95% CI: 0.80, 0.96] vs bounding box: AUC, 0.85 [95% CI: 0.75, 0.94];