A combined UNet++ and LSTM approach for breast ultrasound image segmentation
Saba Hesaraki, Morteza Akbari, Ramin Mousa, Mehrshad Eisaei, Amir Ali Bengari, Roghayeh Taghavi · Franklin Open · 2025
Breast cancer is one of the most common causes of death among women worldwide. This underscores the critical importance of early detection and accurate segmentation of tumors. Deep learning techniques, particularly UNet and UNet++ architectures, have shown remarkable performance in segmenting breast ultrasound images. However, these models do not capture the temporal features in the images. To address this limitation, this study introduces a novel hybrid architecture that combines the encoder–decoder structure of UNet++ with long short-term memory (LSTM) networks to integrate both spatial and temporal features effectively. Additionally, a multi-scale feature extraction module and self-attention mechanisms have been embedded to enable hierarchical feature fusion for capturing richer contextual information. The proposed end-to-end segmentation pipeline was rigorously evaluated using the BUSI with GT dataset; a publicly available dataset of ultrasound-based breast cancer images with corresponding ground truth masks. The model achieved outstanding performance, with an Accuracy of 98.88%, Specificity of 99.53%, Precision of 95.34%, Sensitivity of 91.20%, F1-Score of 93.74%, and Dice coefficient of 92.74%. These results represent a promising improvement over existing methods, suggesting that the proposed approach could serve as a valuable tool within clinical workflows. This study proposes a model called UNet++ + LSTM for the segmentation. As a modified network of Unet++, the model is mainly composed of three parts: a multi-scale feature extraction module, an attention block, and a multi-task learning module. Additionally, This model can process high-frequency local information to improve segmentation.