U-NeTrans at the Edge: Precision and Adaptability in Medical Image Analysis through Segment-based U-Net and Transformer Integration
Saeed Iqbal, Adnan Nabeel Qureshi, Musaed A. Alhussein, Khursheed Aurangzeb, Huihui Helen Wang · Research Square · 2024
Abstract In the domain of edge computing for medical image analyses, utilizing advanced deep learning algorithms has shown promise in boosting precision and adaptability. Medical image analysis frequently requires striking a compromise between a more comprehensive contextual understanding and specific local information. Even though U-Net with Transformers integration is becoming more and more common, issues like patch flattening and magnification sensitivity still exist. By using a unique approach, U-NeTrans directly addresses these problems by segmenting patches instead of flattening them into tokens. This novel method avoids patch flattening issues while maintaining fine-grained local details. Additionally, U-NeTrans has the benefit of allowing for variable patch sizes in a single architecture, which reduces susceptibility to magnification and supports a range of image resolutions. U-NeTrans demonstrates its effectiveness in improving accuracy in a range of medical image analysis tasks by consistently outperforming its competitors in the thorough examination conducted against publicly available datasets. Notably, using the linear correlation between segmentation accuracy and the divergence in U-NeTrans outputs at various scales, it provides a useful confidence metric for evaluating test image complexity, with values of 98.97\% for Accuracy, 98.81\% for Precision, 99.68\% for Sensitivity, 98.73\% for Specificity, and an AUROC of 99.19\%. In circumstances involving peripheral medical image analysis, when computational performance is crucial, this characteristic is very important.