Habitat analysis and other AI technologies for Ki-67 prediction in hepatocellular carcinoma: a multi-center study
Min Zhao1, Chongfeng Duan1, Xiaoming Zhou1
1Department of Radiology, The Affiliated Hospital of Qingdao University, Qingdao, 266000, China.
Objectives:
To develop various artificial intelligence (AI) models including radiomics, deep transfer learning (DTL) and habitat analysis models, as well as a combined model, for the preoperative noninvasive prediction of Ki-67 expression in hepatocellular carcinoma (HCC).
Materials And Methods:
A retrospective analysis was performed on 433 patients with pathologically confirmed HCC from two institutions. According to the postoperative immunohistochemical Ki-67 expression levels, patients were divided into a high Ki-67 expression group (n = 320) and a low Ki-67 expression group (n = 113). They were further split into a training set (n = 349) and a test set (n = 84) in chronological order. Univariable and multivariable logistic regression analyses were conducted to identify independent predictors of Ki-67 expression. Feature extraction was performed from radiomics, DTL, and habitat analysis, followed by subsequent model construction. Subsequently, the extracted multiple features were combined with clinical variables to establish a combined model, and a multivariable logistic regression model was used to construct a nomogram for this combined model. The performance of different models was compared using the area under the receiver operating characteristic curve (AUC), along with other diagnostic metrics including sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Statistical comparisons were conducted to evaluate the differences in performance between the models.
Results:
In the test set, the AUC values of the clinical, radiomics, DTL, habitat analysis and combined models were 0.741(95% CI: 0.616-0.866), 0.754(95% CI: 0.633-0.875), 0.805(95% CI: 0.700-0.911), 0.814(95% CI: 0.712-0.917) and 0.819(95% CI: 0.719-0.920), respectively. Among all the models, the combined model achieved the highest AUC value, followed by the habitat analysis model, but there was no statistical significance between the two. In addition, the combined model had the highest accuracy (94.3%) and sensitivity (95.2%) in the training set, while the habitat analysis model showed the highest accuracy (86.9%) and sensitivity (91.4%) in the test set.
Conclusion:
Radiomics, DTL and habitat analysis models based on Gd-EOB-DTPA-enhanced MRI are effective for preoperative noninvasive prediction of Ki-67 expression in HCC, and the combined model nomogram has potential clinical application value.
