Lab-SAM: SAM-based medical image enhancement algorithm for assisted diagnosis

The Segment Anything Model (SAM) is a large-scale visual model developed for universal segmentation tasks. Due to the significant semantic differences between medical images and natural images, SAM is usually unable to provide high-quality segmentation results for medical images directly. Consequently, the investigation of the potential of SAM for medical image analysis represents a significant research topic. While SAM is currently unable to accurately segment target regions in medical images, the generated segmentation masks demonstrate the preliminary identification and localization of tissues and lesions of interest to radiologists in medical images. This implies that the segmentation output of SAM still has a potential value. Therefore, a medical image enhancement algorithm based on the output of the SAM model, Lab-SAM, is proposed. This algorithm utilizes the binary mask generated by the automatic segmentation of SAM and the corresponding information to perform feature enhancement based on the principle of attention on the input image in Lab space, thereby improving the diagnostic accuracy. In order to verify the effectiveness of the algorithm, classification experiments are conducted on different types of medical image datasets using ResNet and MobileNet. The experimental results demonstrate that Lab-SAM is capable of enhancing the training data, thereby facilitating the classification model’s ability to enhance classification accuracy. This provides a viable approach and concept for the application of SAM in disease diagnosis.