Enhancing Breast Masses Detection and Segmentation: A Novel U-Net-Based Approach
Noura Bentaher, Younes Kabbadj, Mohamed Ben Salah · 2023
Breast cancer is a frequent form of cancer for women, and globally it is a major cause of death. Early detection is crucial for effective diagnosis and treatment, and can also help to lower mortality rates. In recent years, Convolutional Neural Networks have been applied to mammography to assist radiologists improve their efficiency and precision. U-Net is a fully convolutional neural network and a popular model that uses the encoder-decoder architecture, it is known for its capacity to effectively segment biomedical images. This work aims to present a new model for breast cancer detection and segmentation. Our first purpose was to discover a reliable segmentation model for breast masses in mammography images by combining the ResNet architecture and U-Net. Subsequently, we trained several standard models for classification in normal and abnormal images, such as ResNet101, ResNet152, DenseNet169, MobileNet, InceptionV3, and NasNet. Furthermore, based on the segmentation model that yielded good experimental results, we created our classification model by modifying the output layers in order to generate a prediction label (Normal or Abnormal) instead of generating a mask. Our study contributes to the field of breast cancer detection by employing a novel approach that performs detection and segmentation directly on the entire mammography image, unlike traditional methods that operate on image patches. The main advantage of the proposed approach is to simplify the workflow of breast masses segmentation by eliminating intermediate steps which can introduce errors and increase the complexity of the system. Furthermore, our approach focuses on reducing false positives, which are a common challenge in breast cancer detection. Our classification model reached better results compared to other experimented classification models. We used the INbreast dataset of mammography images to train and evaluate our models. In addition to that we employed GradCAM in order to visualize the prediction results and to have an interpretation of the results.