Enhancing breast cancer diagnosis: non-invasive prediction of MKI-67 (Ki67) expression using ultrasound images.
Authors
Affiliations (5)
Affiliations (5)
- Faculty of Applied Sciences, Macao Polytechnic University, de Luís Gonzaga Gomes, Macao, 999078, P. R. China.
- Department of Radiation Oncology, Affiliated Hospital (Clinical College) of Xiangnan University, Renmin West No.31, Chenzhou, 423000, P. R. China.
- Department of Physical Examination, Centre for Disease Control and Prevention of Beihu District, Dongfeng No. 6, Chenzhou, 423000, P. R. China.
- School of Medical Imaging, Laboratory Science and Rehabilitation, Xiangnan University, Renmin West No.31, Chenzhou, 423000, P. R. China.
- Faculty of Applied Sciences, Macao Polytechnic University, de Luís Gonzaga Gomes, Macao, 999078, P. R. China. [email protected].
Abstract
This study explores the non-invasive prediction of MKI-67 (Ki67) expression status in breast cancer using preoperative ultrasound image heterogeneity. Data from 432 patients (training set) and 109 (test set) across two medical institutions were analyzed. Tumor regions were automatically outlined using the Swin-unet network, and habitat clustering within these regions was performed using the k-means method. Radiomics and deep learning features (ResNet-101) were extracted from both global tumor regions and habitat subregions. Laboratory data were integrated, followed by the Least Absolute Shrinkage and Selection Operator (LASSO) feature reduction and machine learning modeling to predict Ki67 expression status. Model performance was evaluated using accuracy (Acc), area under the curve (AUC) with 95% confidence intervals (CI), sensitivity (Sen), specificity (Spe), positive predictive value (PPV), negative predictive value (NPV), calibration curves, confusion matrices, and decision curves. The DeLong test was used to compare the diagnostic performance of the composite model with individual models. The results showed that the combined model (Habitat + Global + Laboratory + Deep Learning) achieved the best predictive performance, with Acc, AUC, Sen, Spe, PPV, and NPV of 0.798, 0.838, 0.780, 0.809, 0.711, and 0.859, respectively, in the test set. Calibration curves and confusion matrices confirmed the model's robustness, while decision curves demonstrated its clinical utility. The DeLong test confirmed the composite model's significantly superior AUC compared to several individual models, though not all combined models showed significant differences. However, despite not showing significant advantages in comparisons with some combined models, the composite model, leveraging its unique strength of comprehensively integrating multi-dimensional features, has demonstrated stronger adaptability and stability in real-world clinical application scenarios, providing more reliable support for accurate prediction. In conclusion, preoperative ultrasound image heterogeneity, through the integration of habitat subregion, global tumor, laboratory, and deep learning features, provides valuable insights for predicting Ki67 expression status in breast cancer, enhancing routine preoperative ultrasonography and offering a potential non-invasive method for preoperative Ki67 prediction.