Comparative Analysis of UNET-Based and Transformer-Based Medical Image Segmentation Models on Lungs X-rays

Abstract

In this Research we did to a comparative analysis of UNET based segmentation models and Transformers based Segmentation model on Chest X ray images in medical imaging domain. The parameters for study were Jaccard Index (IoU) Foreground accuracy, Inference time, and Model size. The Hyper parameters such as Augmentations, learning rate, batch size and image size were kept similar. We used three augmentations, batch size of 32, and image size of 256x256. The experimentation environment for all models was Google Collab Pro, and for training Transformers Hugging face was used for loading dataset, models and Fine tuning. The training dataset was used as 80% training, 10% validation and 10% testing. The Jaccard Index (IOU) for UNET was 92.7 and foreground accuracy was 94. The Jaccard Index (IOU) for U2NET was 94 and foreground accuracy was 95. The Jaccard Index (IOU) for Segformer was 97 and foreground accuracy was 97.9. The Jaccard Index (IOU) for DPT was 97 and foreground accuracy was 97.8. The transformers beat accuracies of UNET based models and also performed better in Inference, and model size was insignificant and did not have any effect on performance. Overall performance of Transformers in same environment was better than UNET based models, and we recommend using transformers for medical image segmentation tasks.