Development and Assessment of a Predictive Model for Ki-67 Expression Using Ultrasound Indicators and Non-Morphological Magnetic Resonance Imaging Parameters Before Breast Cancer Therapy

To formulate a predictive model for assessing Ki-67 expression in breast cancer by integrating pre-treatment ultrasound features with non-morphological magnetic resonance imaging (MRI) parameters, encompassing functional and hemodynamic indicators. A retrospective study was conducted on 167 patients. All patients underwent a breast mass biopsy for histopathological and Ki-67 analysis prior to neoadjuvant chemotherapy (NAC) treatment. Additionally, all patients underwent ultrasonography and MRI examinations prior to the biopsy. The recorded variables were Ki-67, apparent diffusion coefficient (ADC) values, Max Slope, time to peak (TTP), signal enhancement ratio (SER), early enhancement rate (EER), time-signal intensity curve (TIC), tumor maximum diameter, tumor margins and boundaries, aspect ratio, microcalcification, color Doppler flow imaging grading, resistance index (RI), and axillary lymph node metastasis. Statistical analysis was performed using the R software package. Normally distributed continuous data are presented as mean ± standard deviation (SD), skewed continuous data as median, and categorical variables as frequency or percentage. The dataset was randomly divided into a modeling group and a validation group following a 7:3 ratio, employing a predetermined random seed. The selection of variables was conducted using the random forest algorithm. Specifically, in the initial analysis, we trained a random forest model using all available variables. By evaluating the Gini importance scores of each variable, we identified those that contributed the most to predicting Ki-67 expression. The predictive model for Ki-67 expression was constructed using selected variables: Maximum Diameter, ADC value, SER value, Max Slope value, TTP value, and EER value. Within the validation group, the evaluation metrics demonstrated an Area under the curve of 0.961 with a 95% confidence interval ranging from 0.865 to 0.995. The model achieved a kappa score of 1.00, precision of 0.949, recall of 1, an F1 score of 0.974, sensitivity of 100%, specificity of 85.71%, a positive predictive value of 94.87%, and a negative predictive value of 100%. The combination of non-morphological MRI parameters and pre-treatment ultrasound features in a breast cancer prediction model powered by RF machine learning demonstrated favorable clinical outcomes and improved diagnostic performance.