MRI-based deep learning model for prediction of Ki-67 expression level in intrahepatic cholangiocarcinoma

Ki-67 expression is a valuable prognostic indicator in intrahepatic cholangiocarcinoma (iCCA). This study aimed to develop a non-invasive Deep learning (DL) model based on magnetic resonance imaging (MRI) and clinical data to predict Ki-67 expression preoperatively.

Patients with iCCA between January 2020 and January 2024 were retrospectively reviewed. Based on immunohistochemistry, patients were classified into high-expression group (Ki-67 > 50%) and low-expression group (Ki-67 ≤ 50%). After image acquisition, tumor segmentation, and preprocessing, a 3D-ResNet10 network was utilized for MRI feature extraction. A feed-forward neural network (FFN) was used for clinical feature extraction. Subsequently, the extracted image and clinical features were fused and fed into a lightweight classifier to perform Ki-67 level prediction. The performance of the DL models was evaluated via 5-fold cross-validation using accuracy, precision, sensitivity, specificity, F1-score, and the area under the curve (AUC).

A total of 117 patients were ultimately included (30 cases in the high-expression group and 87 cases in the low-expression group). The model based solely on clinical variables demonstrated limited predictive efficacy (mean AUC: 53.18 ± 14.60%). The model utilizing MRI features alone exhibited high specificity (86.14 ± 10.25%) but modest sensitivity. In contrast, the fused DL model significantly outperformed both individual models, attaining a mean AUC of 84.24 ± 8.82%, an accuracy of 87.17 ± 5.95%, a specificity of 93.07 ± 9.35%, and a markedly improved sensitivity of 70.00 ± 19.44%, yielding a robust F1-score of 73.42 ± 12.28%.

We developed and validated a non-invasive DL model integrating contrast-enhanced MRI and clinical parameters that accurately predicts Ki-67 expression in iCCA, demonstrating superior diagnostic performance.