Optimizing BIRADS Classification for Breast Cancer in Moroccan Women: Comparing Four Transfer Learning Models

Abstract

Breast cancer has a significant impact on women's mortality rates in Morocco. Currently, BIRADS Scoring system is considered the standard for diagnosing breast cancer using the correct classification of mammographic images. This research explores the use of four pre-trained deep learning models: NASNet, VGG16, VGG19, and ResNet to automatically categorize BIRADS using transfer learning. Our methodology includes comprehensive preprocessing and fine-tuning of these models to adapt them to BIRADS categories. The dataset used consists of mammographic images that were taken from Moroccon women and given labels according to BIRADS categories. To come up with materials for the model training, there is a detailed preprocessing pipeline that includes image resizing, normalization, and augmentation to prepare the data. In the wording of the transfer learning approach, The transfer learning approach involves fine-tuning the selected models to adapt them and to make them fit well with the specific BIRADS classification task. The results show significant improvements in classification accuracy, area under the curve (AUC), and loss, with VGG16 achieving the best performance at 92% accuracy. The study emphasizes the promise of leveraging advanced deep learning techniques to support radiologists and improve clinical outcomes for future research in breast cancer diagnostics, contributing to the growing body of literature on AI-driven solutions in healthcare.