Machine learning algorithm for classification of Ductal carcinoma in situ and minimal invasive breast cancer

Introduction: Breast cancer nowadays is the second common cancer in the world and the most common cancer among women, excluding nonmelanoma skin cancers. Breast cancer is not just one disease, it has different types and subtypes that depend on the affected specific cell in the breast. Cancer can be classified into two types according to whether it has spread: non-invasive and invasive breast cancer. The most frequent kind of non-invasive breast cancer is ductal carcinoma in situ (DCIS). DCIS is cancer that starts in a duct and has not spread into any surrounding breast tissue. Some DCIS patients will not develop the invasive disease, and this has been suggested as a risk of screening mammography. Breast cancers that are invasive have grown outside of the ducts or lobules into the surrounding tissue. As size of the tumor decreases, patients with invasive breast cancer have a better chance of surviving. Despite the prognostic factors, a small percentage of patients with invasive tumors of 10 mm or less (T1a and T1b) die from their cancer. Many studies have been conducted examining traditional histopathological characteristics, including lymph node status, tumor size, histological grade, margin width, and many other biological markers of prognosis. The use of these prognostic factors, while appealing in principle and effective in larger tumors, presents challenges in small tumors. The identification of breast cancer types at an early stage enables patients to choose less invasive treatment options. The purpose of our study was to develop a machine-learning classification model to differentiate DCIS and minimal invasive cancer using clinical characteristics, mammography findings, ultrasound findings and histopathology features.
Method: Clinical data, mammography findings and ultrasound findings of 420 biopsy-confirmed breast cancer cases were analyzed retrospectively to diagnose DCIS and minimal invasive cancer. The subtypes were categorized based on the histopathology and size of lesion on histological assessment. Four groups of features including clinical data, mammography findings, ultrasound findings and histology findings are used for classification by machine learning. The machine learning techniques used in this study include XGboost, Random Forest, Single Vector Machine, Gaussian Naive Bayes, K-Nearest Neighbor, and Decision Tree Classifier. To classify two types of breast cancer, we mainly focus on the XGBoost algorithm trained on clinical characteristics, mammography (MMG) findings, ultrasound (US) findings, and histopathology features that are associated with DCIS and minimal invasive breast cancers. The study used the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, accuracy, precision, and F1 score as measures of model performance. Additionally, this research determined the importance of features by using XGboost and SHapley Additive Explanations (SHAP).
Results: The results of this model were validated in 378 women and tested in 42 women (mean age, 58.8 years ± 12.2). The model has high classified performance when combining features importance, with the highest accuracy reaching 0.84 (95% confidence interval [CI]: 0.77, 0.90), an AUC of 0.93 (95% CI: 0.86, 0.96), with the specificity of 0.73 (95% CI: 0.64, 0.82) and sensitivity of 0.91 (95% CI: 0.73, 0.95). The five most important features illustrated by XGBoost were the presence of calcification on MMG, the existence of lymph node, the presence of microcalcification on histopathology, the shape of the mass on US image, and the orientation of mass on US image, and the orientation of mass on US image.
Conclusion: XGBoost model combining clinical characteristics, mammography findings, ultrasound findings, and histopathology features, can be applied to classify breast cancer at a level equivalent to radiologists and has the potential to detect early invasive breast cancer.