深度学习模型无创预测乳腺癌Ki-67表达及预后分层：多中心回顾性研究

OBJECTIVE: Ki-67 correlates with prognosis for patients with breast cancer. However, the evaluation of Ki-67expression relies on pathological analysis and invasive biopsy, which hinders its wide adoption. This work sought to develop a noninvasive Ki-67 prediction model for breast cancer patients through ultrasound and clinical information and evaluate model performance in risk stratification of lymph metastasis and prognosis.

MATERIALS AND METHODS: Clinical, ultrasound, pathological, and prognostic information were collected from four centers to develop a deep learning (DL) model. Ultrasound features were extracted by ResNet-50 and integrated with clinical information through logistic regression. Class activation mapping and nomograms were used to visualize the prediction process. Area under curve (AUC), confusion matrices, calibration curves, and decision curve analysis were used to evaluate model performance on Ki-67. Prognostic relevance was evaluated with lymph node metastasis and recurrence-free survival (RFS).

RESULTS: From January 2021 to December 2024, 456 patients from three centers were collected as training (n = 264), validation (n = 96), and internal test (n = 96) sets, 204 patients from an independent center were collected as an external test set. In the external set, the combined model achieved satisfactory performance on Ki-67 (AUC = 0.828, 95% CI: 0.761-0.890). High Ki-67 group showed higher lymph metastasis rates (67.7% vs 16.2%, p < 0.001) and worse RFS (p = 0.041) than the low Ki-67 group. The combined model achieved the best predictive ability on recurrence in the first 6 months after operation (AUC = 0.820).

CONCLUSION: This noninvasive model could predict Ki-67 status, classify the risk of lymph metastasis, and provide prognostic insights. Its wide application would contribute to the formulation of individualized treatment plans and follow-up strategies.

KEY POINTS: Question Ki-67 is an important pathological information in breast cancer and is meaningful for lymph metastasis and prognosis. However, it can currently only be evaluated through invasive testing. Findings We developed a non-invasive model based on ultrasound and clinical information, which can predict Ki-67 status (accuracy = 0.828), classify lymph metastasis (accuracy = 0.765), and provide prognostic insights (p = 0.041). Clinical relevance This model enabled preoperative prediction of Ki-67 status and lymph metastasis for patients with breast cancer, thereby informing surgical planning. Furthermore, it demonstrated prognostic utility, facilitating the development of personalized patient follow-up strategies.