Preoperative prediction of Ki-67 expression in hepatocellular carcinoma using multi-sequence MRI radiomics: a non-contrast approach integrating T2-weighted, diffusion-weighted, and apparent diffusion coefficient imaging.

Authors

Abstract

Ki-67 is a key immunohistochemical marker of tumor proliferative activity in hepatocellular carcinoma (HCC); however, non-invasive assessment of Ki-67 status before surgery remains an unmet clinical need. This study aimed to develop and internally validate multi-sequence magnetic resonance imaging (MRI) radiomics models based exclusively on non-contrast sequences for preoperative prediction of Ki-67 expression. This retrospective study included 91 patients with 97 pathologically confirmed HCC lesions ( high Ki-67, ≥ 10%, n = 55; low Ki-67, < 10%, n = 42). Three-dimensional volumes of interest were manually segmented on T2-weighted imaging ( T2WI), diffusion-weighted imaging ( DWI), and apparent diffusion coefficient ( ADC) maps, and 5055 radiomic features ( approximately 1684 per sequence) were extracted were extracted from the three sequences. After feature selection with Pearson correlation coefficient filtering and the Kruskal-Wallis test, three multilayer perceptron ( MLP) models were constructed: Model-T ( T2WI), Model-D ( DWI + ADC), and Model-M ( combined T2WI, DWI, and ADC). The models were developed with five-fold cross-validation in the training cohort and were evaluated in a held-out testing cohort. This was a single-center retrospective study, and validation was internal only. Model performance was assessed using the area under the receiver operating characteristic curve ( AUC), sensitivity, specificity, and accuracy. No significant differences in clinical characteristics were observed between the high and low Ki-67 groups. In the training cohort, the apparent (resubstitution) AUCs of Model-T, Model-D, and Model-M were 0.715, 0.823, and 0.860, respectively (representing model fit rather than cross-validated performance). In the held-out testing cohort, the corresponding AUCs were 0.707, 0.692, and 0.716. In the testing cohort, Model-M achieved a sensitivity of 68.8%, a specificity of 53.8%, and an accuracy of 62.1%, whereas Model-D showed the most balanced performance (sensitivity, 68.8%; specificity, 76.9%; accuracy, 72.4%). The dominant features were texture features derived predominantly from the ADC sequence. The cross-validated training AUCs were considerably lower (range, 0.522-0.585). In this single-center, internally validated study, non-contrast multi-sequence MRI radiomics models integrating T2WI, DWI, and ADC showed moderate discrimination of Ki-67 expression in HCC. The contrast-free strategy is potentially applicable to patients with contraindications to gadolinium and warrants external validation; These findings represent internal validation only and require confirmation in external, preferably prospective multicenter, cohorts before any clinical application can be considered.

Tags

Topics