Enhancing breast tumor segmentation in ultrasound images with EfficientNet-B4 and UNet++ integration

This is the authors' abstract. We don't add key points for this paper.

Purpose:: As breast cancer remains one of the leading causes of mortality among women, it is crucial to establish accurate diagnosis and staging as well as treatment strategies using rapid and accurate detection of lesions. In numerous clinical applications, ultrasound is preferred due to a lack of radiation exposure and cost-effectiveness; nonetheless, the imaging modalities’ inherent speckle noise, low contrast levels, and varied morphologies of tumors limit the use of conventional image segmentation techniques. Methods:: This study employs a deep learning model that incorporates an EfficientNet-B4 encoder with a UNet++ decoder to improve performance over prior studies by incorporating both compound-scaled feature extraction (encoder) and dense multi-scale skip connection architecture (decoder) to overcome the limitations listed above. A composite loss integrating binary cross-entropy and Tversky loss is utilized to address class imbalance and enhance boundary precision. Results:: On the reserved test set, the model achieves a pixel-wise accuracy of 98.68% and a Dice score of 0.9599, surpassing previous state-of-the-art designs by up to 3% in overlap metrics while sustaining a tiny parameter footprint of 31 million. Conclusion:: These results indicate that the model could be a reliable and fast way to use ultrasound to help diagnose breast cancers in real time.