Evaluating Various Training Strategies and Different Backbones for U-Net in Breast Mass Segmentation

Abstract

Breast cancer is a common type of cancer which affects women globally, being a significant contributor to mortality rates. Early detection is crucial in the treatment of breast cancer, ultimately reducing mortality. This study seeks to evaluate the effectiveness of transfer learning in improving breast mass segmentation using U-Net, a fully convolutional neural network that is renowned for its robustness in biomedical images segmentation. Its effectiveness lies in the use of and encoder-decoder architecture to capture coarse and fine details and skip connections as an attention mechanism. In this paper we leverage and compare different models such as ResNet, DenseNet, MobileNet, Inception, and EfficientNet, accessible in Keras, to serve as the encoder component of U-Net. The segmentation models were trained and evaluated on the INbreast mammographic dataset, comprising original and annotated mammography images. The methodology involved either doing all the training or employing transfer learning. Both segmentation models were trained using two loss functions, the Jaccard distance, and the Dice Loss. Results indicate that both U-Net and ResNet-Unet are particularly suited for breast mass segmentation and that transfer learning is unreliable when used in the specific domain of mass segmentation.