Abstract
Introduction. Histological and molecular analysis of breast tissue is essential for the diagnosis, prognosis, and treatment of breast cancer. Key biomarkers include progesterone and estrogen receptors, as well as the human epidermal growth factor receptor 2 (HER2). HER2 overexpression indicates an aggressive subtype of breast cancer but enables targeted therapies that improve survival rates. However, its evaluation faces challenges, ranging from sample quality to interpretation variability. The College of American Pathologists classifies HER2 overexpression into four categories, but variations around the 10% expression threshold can lead to misinterpretations. Objective. To present an automated technique for classifying HER2-overexpressing cells in histological slides. Materials and methods. The Cross-Industry Standard Process for Data Mining (CRISP-DM) methodology was applied using samples of 89 patients from the Unidad de Diagnóstico en Patología, covering all four HER2 expression levels. Deep learning techniques were employed, leveraging neural networks and vision transformer models through transfer learning. Additionally, a usability evaluation was conducted on the final version of the software. Results. The ViT-B/16 model achieved a classification accuracy of 90,65%, while the tool was evaluated with an acceptable level of satisfaction in its clinical application. Conclusion. Artificial intelligence demonstrated high accuracy and consistency in HER2 classification, reducing diagnostic variability and improving objectivity. However, further optimization of processing efficiency is required for broader applicability.