Are 2D MRI slices equally important in microvascular invasion prediction: a study based on multiple instance learning with attention.

backgroundMicrovascular invasion (MVI) is a critical prognostic indicator of hepatocellular carcinoma (HCC). Although MRI-based prediction of MVI has drawn great attention, many previous studies focused on either 3D volumetric regions of tumors or the slices of the largest tumors size in the axial plane. This study aimed to investigate the predictive value of 2D MRI slices on MVI, based on a multiple instance learning (MIL) method with an attention block to estimate the slice-wise importance.

methodsA total of 174 patients (training cohort: 121 patients; testing cohort: 53 patients) were evaluated in this retrospective study. MRI scans of arterial phase, venous phase and delayed phase, as well as sixteen clinical characteristics were collected. A transformer-based model was utilized to extract features from the original and dilated tumor regions (with 5 mm and 10 mm margins). A MIL based method was developed to predict MVI, with an attention block to estimate the slice-wise weights. Eighteen phase-specific and six combined-phase models were trained and tested. Three hybrid models were constructed from the best-performed models. Area under the curve (AUC), precision, accuracy, F1 were utilized to evaluate model performance. Centered kernel alignment (CKA) analysis was employed to assess the value of the learned weights of MRI slices by estimating their correlation with histopathological image.

resultsArterial phase outperformed other phase-specific and combined-phase models on test AUC (two best models: 0.78, 95%CI: 0.63, 0.91; 0.78, 95%CI: 0.64, 0.90). The hybrid model using soft voting strategy achieved the highest test AUC (0.80, 95%CI: 0.69, 0.92) in this study. CKA analysis of one patient’s data showed that the top 5 MRI slices of the highest weights have stronger correlation with MVI than others (0.87 versus 0.75).

conclusionThe group of 2D MRI slices demonstrated predictive value for MVI prediction using a MIL method. The learned weights of MRI slices from the attention block indicate their potential relevance with MVI. These findings may help to identify the most relevant MRI slices for MVI prediction, which may be utilized in MVI diagnosis in future applications. CLINICAL TRIAL NUMBER: Not applicable.