Enhanced Segmentation of Lesions in Medical Images using DTA-Transunet

The aim of the study is to improve lesion segmentation from medical images using the DTA-TransUNet model, which includes Triplet Attention and DCNv3 to improve flexibility and accuracy. Its performance is verified and compared with the CNN model to ensure its ability to perform accurate segmentation. The study for Group 1 (Comparison Group) The CNN model applied on 50 randomly chosen images of the ISIC 2017 dataset and Group 2 (Intervention Group) The TransUNet model used for the same 50 images. The two groups were trained with the same hyperparameters in MATLAB with a learning rate of 0.001, a batch size of 16, and 50 epochs. The G Power of 80%, threshold of 0.05, and 95% confidence interval were employed. Performance evaluation was performed in terms of the Dice Similarity Coefficient, Intersection over Union, and Jaccard Index, showing superiority of TransUNet over CNN in segmentation precision. The DTA-TransUNet model yielded DSC = 86.2%, IoU = 78.4%, and Jaccard Index = 79.1% compared with the TransUNet (DSC = 83.8%) and U-Net (DSC = 79.5%) with statistical significance at p < 0.05. It is found that the DTA-TransUNet model with Triplet Attention and DCNv3 outperforms the traditional U-Net and TransUNet models.