An segformer-mix vision transformer-based GIST segmentation in fused CT-PET images

Accurate segmentation of gastrointestinal stromal tumour (GIST) is always challenging task due to tissues intestines may have similar intensity values, which makes boundary delineation harder with only one source of contrast. Presently, clinicians make use of CT scans for manually annotate and diagnosis of GIST, which is ineffective and susceptible to subjectivity. To improve tumour delineation and effective segmentation of GIST, here we proposed two stage deep learning model for this purpose. In the first stage an DIF-Net is used to fuse CT and PET images, helps to improve quality of image for accurate delineation. In further stage the tumour region is effectively segmented by employing a Segformer architecture supplemented with a Mix Vision Transformer (MViT) encoder. The proposed methodology combines the powerful fusion characteristics of DIF-Net with the efficient segmentation characteristics of Segformer, allowing correctly capturing both local and global features, ultimately achieving superior segmentation results. The superiority of model performance is evaluated with multiple performance parameters for segmentation like mean Intersection Over Union (IoU), Pixel Accuracy, and dice coefficient, and for fusion it is PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural Similarity Index), MI (Mutual Information), UIQI (Universal Image Quality Index) and Edge Preservation. Experimental results show that the proposed method achieves better segmentation and fusion quality than existing models, enhancing the accurate diagnosis.