Prostate Segmentation in MRI Using Transformer Encoder and Decoder Framework

To develop an accurate segmentation model for the prostate and lesion area to help clinicians diagnose diseases, we propose a multi-encoder and decoder segmentation network, denoted Muled-Net, which can concurrently segment the prostate and lesion regions in an image. The model performs parallel calculations for dual input. In two encoder branches of the model, a new transformer encoder is used to overcome the fact that only information from the neighborhood pixels can be captured, increasing the ability to capture global dependencies. Furthermore, given the usually small size of the lesion, ASPP and feature fusion are merged to expand the perceptual field and retain more contextual information of the shallow layer in decoder. To the best of our limited knowledge, there is no public dataset for the segmentation of the prostate and its lesion regions. So we made a publicly usable dataset. Muled-Net is compared with other deep learning methods, FCN, U-Net, U-Net++, and ResU-Net with four-fold cross-validation. Of all 218 subjects, 140 healthy individuals and 78 patients with prostate cancer were included in this work. Average Dice of 95%, Iou of 89%, sensitivity of 94%, 95HD of 9.56, and MSD of 0.66 are achieved for the prostate segmentation and average Dice of 89%, Iou of 82%, sensitivity of 92%, 95HD of 11.16, and MSD of 1.09 for the segmentation of the prostate lesion regions. The performance of the proposed model has made significant improvements to the segmentation of the lesion regions in particular, suggesting that the model could be considered as an auxiliary tool to ease the workload of physicians and help them in making treatment decisions.