EU2-Net: A Parameter Efficient Ensemble Model With Attention-Aided Triple Feature Fusion for Tumor Segmentation in Breast Ultrasound Images

Breast cancer is a widespread and serious health concern, emphasizing the utmost importance of accurate screening and diagnosis of breast tumors. The ultrasound imaging has emerged as a widely adopted method for detecting breast tumors due to its versatility and accessibility. The segmentation of breast ultrasound (BUS) images plays a crucial role in the identification and localization of breast tumors, forming an essential component of computer-aided diagnosis (CAD) systems. However, the complexity of BUS images, characterized by intricate patterns, closely resembling intensity distributions, varying tumor shapes, and unclear boundaries, presents significant obstacles in achieving accurate segmentation of breast lesions. In this article, we introduce EU2-Net, a novel ensemble model that extends the capabilities of the U2-Net architecture, a standard framework used for image segmentation. EU2-Net is a lightweight ensemble architecture that replaces the conventional convolutional layers with separable convolutions. This architectural refinement significantly reduces the number of trainable parameters, enhancing the model’s training efficiency. Unlike the conventional ensemble techniques that require multiple models and a parameter explosion, we introduce a weighted averaging ensemble mechanism with learnable weights, seamlessly integrated within the U2-Net architecture. To enhance the model performance further, we introduce an attention-aided triple feature fusion (A2TF2) technique. This enhances both encoder and decoder features using feature similarity-based attention and squeeze-and-excitation (SE) channel attention. In addition, we incorporate edge features extracted by the Sobel filter from the encoder features, enriching the decoder with boundary-aware information. These combined multidomain features provide the decoder with a comprehensive set of generic and detailed features, thereby improving segmentation accuracy. We demonstrate the effectiveness of our method on two publicly accessible datasets: BUSI and UDIAT and achieve state-of-the-art (SOTA) results. The code will be made available at: https://github.com/AyushRoy2001/EU-2net.