Intratumoral and peritumoral radiomics model based on abdominal ultrasound for predicting Ki-67 expression in patients with hepatocellular cancer

Abstract

Background: Hepatocellular cancer (HCC) is one of the most common tumors worldwide, and Ki-67 is highly important in the assessment of HCC. Our study aimed to evaluate the value of ultrasound radiomics based on intratumoral and peritumoral tissues in predicting Ki-67 expression levels in patients with HCC.

Methods: We conducted a retrospective analysis of ultrasonic and clinical data from 118 patients diagnosed with HCC through histopathological examination of surgical specimens in our hospital between September 2019 and January 2023. Radiomics features were extracted from ultrasound images of both intratumoral and peritumoral regions. To select the optimal features, we utilized the t-test and the least absolute shrinkage and selection operator (LASSO). We compared the area under the curve (AUC) values to determine the most effective modeling method. Subsequently, we developed four models: the intratumoral model, the peritumoral model, combined model #1, and combined model #2.

Results: Of the 118 patients, 64 were confirmed to have high Ki-67 expression while 54 were confirmed to have low Ki-67 expression. The AUC of the intratumoral model was 0.796 (0.649-0.942), and the AUC of the peritumoral model was 0.772 (0.619-0.926). Furthermore, combined model#1 yielded an AUC of 0.870 (0.751-0.989), and the AUC of combined model#2 was 0.762 (0.605-0.918). Among these models, combined model#1 showed the best performance in terms of AUC, accuracy, F1-score, and decision curve analysis (DCA).

Conclusion: We presented an ultrasound radiomics model that utilizes both intratumoral and peritumoral tissue information to accurately predict Ki-67 expression in HCC patients. We believe that incorporating both regions in a proper manner can enhance the diagnostic performance of the prediction model. Nevertheless, it is not sufficient to include both regions in the region of interest (ROI) without careful consideration.

Keywords: Ki-67 Antigen; computer assisted diagnosis; hepatocellular cancer; machine learning; radiomics; ultrasonography.

Figure 1

Flowchart of the inclusions and exclusions criteria of participants.

Figure 2

Representative immunohistochemistry Ki-67 staining patterns and dot plots assessing the percentage of Ki-67 staining cells (original magnification, 200x). (A) Low Ki-67 expression (7%); (B) High Ki-67 expression (80%). Brown-stained nuclei were considered positive Ki-67 expression.

Figure 3

An example of delineating region of interest (ROI) on abdominal ultrasound imaging in ITK-SNAP software. (A): original image; (B): intratumoral ROI. (C): peritumoral ROI; (D): combined ROI (intratumoral tissue + peritumoral tissue).

Figure 4

Receiver operating characteristic (ROC) curve results from different modeling methods in the intratumoral model and peritumoral model. The SVM algorithm showed the highest diagnostic performance with an AUC value of 0.796 (0.649-0.942) in the intratumoral model (A) and 0.772 (0.619-0.926) in the peritumoral model (B).

Figure 5

The waterfall plot displays the model’s performance in the intratumoral model validation group (A) and the peritumoral model validation group (B). The height of each bar in the chart represented the model predicted value minus the cut-off. The bars above the y=0 line indicate that the model predicts high Ki-67 expression, while the bars below the y=0 line indicate that the model predicts low Ki-67 expression.

Figure 6

(A) Receiver operating characteristic curve analysis of combined model#1; (B) Receiver operating characteristic (ROC) curve results of different modeling methods in combined model#2. The LR algorithm showed the highest diagnostic performance with an AUC value of 0.762 (0.605-0.918).

Figure 7

The waterfall plot displays the model’s performance in combined model#1 validation group (A) and combined model#2 validation group (B). The height of each bar in the chart represented the model predicted value minus the cut-off. The bars above the y=0 line indicate that the model predicts high Ki-67 expression, while the bars below the y=0 line indicate that the model predicts low Ki-67 expression.

Figure 8

(A). Comparison of receiver operating characteristic curves for differentiation of the four models; (B). Calibration curve of combined model#1 presented a good agreement between the predicted and pathologically confirmed Ki-67 status with dotted line (actual calibration) closed to dashed line (perfect calibration); (C). Decision curves for the intratumoral model, peritumoral model, combined model#1, and combined model#2.