ADVANCEMENTS IN BREAST CANCER DETECTION: AN ANALYSIS OF PRE-TRAINED CNN MODELS AND TRANSFORMER-BASED TRANSFER LEARNING ON MAMMOGRAM SLIDES

Abstract

Breast cancer remains a formidable disease, claiming millions of lives worldwide annually. Timely and precise detection of malignant tumors is paramount to enhancing patient prognosis. In our investigation, we distinguish between benign and malignant tumors and evaluate the efficacy of fourteen pre-trained convolutional neural network (CNN) models using the INbreast cancer dataset. Leveraging transfer learning methodologies to discern malignant tumors across 410 mammograms. In a pioneering, we conduct a novel analysis of attention weight scores generated by Hierarchical Attention Networks (VIT-L16) within a deep dense transfer network to discern the most influential factors in predicting malignant tumors. Our methodological approach encompasses a range of evaluation metrics for all pre-trained models, laying the groundwork for the development of an automated system proficient in identifying various breast lesions. The integration of transformers with transfer learning presents several benefits. Firstly, transformers excel at capturing long-range dependencies within data, enabling them to effectively analyze intricate patterns and relationships present in medical images, such as those in mammograms. Secondly, by leveraging pre-trained transformer models, we can expedite the training process and mitigate the need for vast amounts of labeled data, thus facilitating the development of accurate diagnostic systems even in scenarios where labeled medical datasets are scarce. Ultimately, the fusion of transformers with transfer learning equips us with powerful tools to enhance the early and accurate detection of breast cancer, with profound implications for the realms of medical imaging and computer-aided diagnosis (CAD). Our experimental findings reveal ResNet50 as the top-performing model, achieving an accuracy of 73.09%.