Step edge and edge loss enhanced Transformer for medical image segmentation

Abstract

Objective

Knee osteoarthritis (KOA) is the most common degenerative joint disease, significantly impacting patients' quality of life. Currently, KOA assessments using ultrasound images rely on complex, time-consuming manual measurements. Developing automatic, accurate medical image segmentation methods for clinical diagnosis is crucial. However, due to challenges like blurred boundaries and irregular shapes in medical images, existing methods show poor edge accuracy and morphological differences.

Methods

To address these issues, we propose IELSEMT model. The model proposes a step edge module, incorporating it into both global and local branches during training. Additionally, after integrating the step edge module, the network's edge feature extraction capability is enhanced. During the network's backpropagation phase, we propose a weighted loss function that incorporates edge loss terms, specifically using the weighted sum of cross-entropy and mean absolute error as the network's loss function, reducing edge region errors in network segmentation results.

Results

Compared to the convolutional baseline, the IELSEMT model improves F1_Score by approximately 5 % and IoU by 15 %. Compared to the Med baseline model, IoU and F1 Score are about 4 % higher. Additionally, by incorporating edge loss terms during the backpropagation phase, the network training speed is enhanced. The IELSEMT model can quickly locate segmentation object edges and trains faster, achieving a convergence speed twice that of the MedT model.

Conclusion

The IELSEMT model proposed in this paper incorporates full-scale and edge detection concepts, introducing edge information into the decoding phase to achieve more accurate edge information acquisition, improving segmentation accuracy with clearer and more precise edges. Unlike other Transformer-based models, our proposed method does not require pre-training on large-scale datasets. Finally, we conducted experiments on knee joint ultrasound datasets, where IELSEMT demonstrated superior performance over both CNN and other related Transformer-based architectures.