TransCS-Net: A hybrid transformer-based privacy-protecting network using compressed sensing for medical image segmentation

Although deep learning models are growingly being employed in medical image segmentation, directly using images to train them may violate patient privacy and data protection laws. To tackle this issue, a novel compressed sensing-based transformer network, namely TransCS-Net, is presented. It first used a compressed sensing (CS) module to compress medical images into low-dimensional measurements, and employed an input projection to adjust their resolution to the same resolution to account for their different resolutions caused by different CS ratios. Then, a dual encoder was adopted to capture richer features by combining long-range dependencies extracted by transformers and local features learned by convolutional neural networks. Extensive experiments showed the proposed TransCS-Net has achieved state-of-the-art performance compared with other models under different CS ratios in five public datasets. Especially our model with a CS ratio of 10% achieved slightly better performance than other methods directly using the original images in the Red Lesion Endoscopy and PH2 datasets. Consequently, the CS has now been employed for the first time in deep learning models that achieves medical image segmentation with competitive results, ensuring patient privacy.