Artificial intelligence in determining the molecular biological subtype of breast cancer; Искусственный интеллект в определении молекулярно-биологического подтипа рака молочной железы

Aim. To investigate the possibility of using radiation diagnostic data to determine various molecular subtypes of breast cancer (BC) using artificial intelligence technologies. Materials and methods. The material for the study was retrospective data of 344 patients treated at the Sverdlovsk Regional Oncology Dispensary in the period from 2021 to 2023. The average age of the study sample was 56.8 ± 10.6 years, ranging from 33 to 82 years. All patients were diagnosed with BC, confirmed histologically. Molecular subtypes of BC were assessed based on trepan biopsy and surgical material. All patients underwent mammographic, ultrasound, and magnetic resonance imaging examinations, and sets of diagnostic features were identified that most accurately correspond to various molecular subtypes of BC. To achieve this goal, the authors identified the following diagnostic features: age, maximum diameter of the formation measured for various methods of radiation diagnostics, morphological features (contours, spatial orientation, shape of the detected formations or areas of reconstruction, heterogeneity of the structure of formations, presence of calcifications, characteristics of blood flow in the tumor) and dynamic parameters of paramagnetic accumulation during magnetic resonance imaging of the mammary gland. Based on the histological examination data, the degree of tumor differentiation (G), proliferative activity index (Ki-67), regional lymph node status (presence or absence of metastases), and molecular-immunohistochemical tumor subtype were assessed. An analysis was conducted to determine whether there was a statistically significant relationship between diagnostic features and molecular subtypes of BC. The analysis was performed by conducting chi-square tests for features and subtypes (classes) of BC, previously converted to binary form. From the arrays of values selected for the study of diagnostic features, training and test samples were formed, and an algorithm for the classification model of artificial intelligence was determined. The accuracy of BC typing was ensured by using a combination of 7 diagnostic features and 6 classification models: five single-class and one multi-class. The gradient boosting algorithm (Gradient Boosting Regressor) was used to train single-class models. The strategy “one (class) versus the rest” was used to train the multi-class model using the One Vs Rest Classifier and gradient boosting (Gradient Boosting Classifier) algorithms. The quality of the trained model was tested on test data. Statistical data processing, development of classification models, their testing and assessment of the quality of training were performed in the Jupyter Notebook environment v.6.5.2. Results. The training quality indicators of single-class models for recognizing BC subtypes were as follows: sensitivity in determining luminal A subtype (LA) was 67.0 %, luminal B subtype (LB) – 72.7 %, luminal B HER2-positive subtype (LBH) – 81.8 %, non-luminal HER2-positive (HER) and triple negative breast cancer (TNC) – 100 %. The specificity was 90.2 % for LA, 83.0 % for LB, 89.7 % for LBH, 98.3 % and 93.5 % in the cases of HER and TNC, respectively. The area under the ROC curve (AUC) depending on the molecular subtype was determined as follows: for LA – 0.88, for LB – 0.86, for LBH – 0.87, for HER – 0.96, and for TNC – 1.000. The multiclass model also showed low sensitivity values, except for the TNC (100 %) and HER (85.7 %) subtypes, low levels of positive predictive value for all subtypes, except for TNC (91.7 %), and high specificity and negative predictive value for all subtypes. The area under the ROC curve for the multiclass model was for the subtypes: LA – 0.88, LB – 0.86, LBH – 0.86, HER – 0.95 and for TNC – 1.00. Conclusion. The possibility of using certain combinations of diagnostic features obtained as a result of radiation diagnostic methods to determine the probability of a molecular biological subtype of BC was proven. This indicates the presence of prerequisites for the creation of a new diagnostic tool for typing BC using classification models of artificial intelligence. In the future, its implementation will reduce the likelihood of an error in determining the molecular biological subtype of BC, especially in situations where the doctor»s opinion and the results of the immunohistochemical study do not coincide. © 2025 Elsevier B.V., All rights reserved.