MULTI-MODAL DEEP LEARNING FRAMEWORK FOR BREAST CANCER DETECTION USING MAMMOGRAM AND HISTOPATHOLOGICAL IMAGES

Abstract

Breast cancer remains one of the leading causes of mortality among women worldwide, necessitating the development of efficient and accurate diagnostic tools. This project presents an advanced deep learning-based diagnostic framework that integrates dual imaging modalities: mammogram and histopathological images, for enhanced breast cancer detection. The system is developed using Python as the backend language, Flask as the web framework, and HTML, CSS, and JavaScript for the frontend interface. For the mammogram analysis, two separate convolutional neural network architectures, Xception and MobileNet, were employed. The mammogram dataset comprises 3,044 training images classified into four BI-RADS categories: BI-RADS 1 (1865 images), BI-RADS 3 (387), BIRADS 4 (408), and BI-RADS 5 (384). The Xception model achieved a training accuracy of 96% and a test accuracy of 90%, while the MobileNet model demonstrated superior generalization with a training accuracy of 96% and a test accuracy of 94%. In parallel, the histopathological image classification module utilizes the DenseNet201 architecture to distinguish between benign and malignant tissue samples using the BreakHis 400X dataset. This dataset includes 1,693 high-resolution biopsy images, with 371 benign and 777 malignant images used for training and 176 benign and 369 malignant images for testing. The DenseNet201 model achieved a training accuracy of 98% and a validation accuracy of 88%. The system provides a comprehensive performance analysis based on Accuracy, Precision, Recall, F1-Score, and Confusion Matrix. MobileNet recorded an accuracy of 94.3%, precision of 81.7%, recall of 94.3%, and F1-score of 94.3%, outperforming the Xception model, which achieved an accuracy of 90.7%, precision of 72.3%, recall of 90.7%, and F1-score of 90.7%. For the histopathological analysis, DenseNet201 achieved an accuracy of 88.1%, precision of 91.3%, recall of 88.1%, and F1-score of 88.1%. Additionally, class distribution statistics are visualized to illustrate the dataset balance and support interpretability. By combining radiological and microscopic imaging modalities, this project offers a robust and intelligent diagnostic solution for breast cancer, enhancing early detection and assisting clinical decision-making with improved accuracy and reliability