BreastCDNet: Breast Cancer Detection Neural Network, Classification and Localization

Abstract

Abstract Breast cancer remains a significant health concern globally, necessitating advanced detection methods for improved patient outcomes. Convolutional neural networks, CNN, have revolutionized object detection by automating the feature extraction process and allowing for the efficient localization and classification of objects within images. BreastCDNet introduces a novel approach to breast cancer detection using CNN. This pioneering CNN method is tailored specifically for the task of ultrasound breast cancer detection and localization, offering a streamlined and effective approach that significantly enhances diagnostic accuracy. Multi-task learning is leveraged by the proposed model, with simultaneous consideration of breast ultrasound image classification and bounding box regression for lesion localization. Intricate features from medical images are extracted by BreastCDNet architecture, facilitating both classification (benign or malignant) and object localization. Separate data generators are employed for classification and bounding box labels to optimize model training. The model's effectiveness is demonstrated by experimental results on the BUSI dataset, where exceptional classification accuracy and precise bounding box predictions are achieved. Key components, including convolutional layers for feature extraction, max-pooling layers for spatial reduction, fully connected layers for predictions, and specific output layers for classification and bounding box regression, are encompassed by the model architecture. The BreastCDNet model is a high-performance breast cancer detection model that achieved 99.14% training accuracy, 97.70% validation accuracy, 0.99 ROC AUC, and 0.97 F1-score on the training set. It is also capable of accurately localizing breast cancer lesions with an IOU score of 0.95.