LiteWaveRep-MedSAM: a lightweight medical image segmentation model based on wavelet transform and reparameterization

As a representative of large-scale general medical image segmentation models, MedSAM's massive parameters incur high computational costs, severely limiting its real-time clinical deployment on mobile and edge devices. To address this challenge, this paper proposes LiteWaveRep-MedSAM-a lightweight, efficient model tailored for mobile deployment. Based on RepViT, it employs macro-level restructuring and incorporates improved wavelet transform and half-convolution techniques to construct the efficient and lightweight visual encoder LiteWaveRepViT. This design significantly reduces computational complexity while ensuring effective extraction of image details and efficient interaction of features. In the decoder, we designed multi-scale channel-adaptive reverse convolutions that achieve high-quality upsampling by solving a regularized least-squares problem in closed form. Thanks to these designs, LiteWaveRep-MedSAM's model parameters have been compressed to6.80Mand computational cost has been reduced to29.72GFLOPs. Experimental results demonstrated that LiteWaveRep-MedSAM, as one of the most lightweight MedSAM architectures to date, exhibits highly competitive performance on multimodal medical image datasets exceeding 100 000 images.The code is available athttps://github.com/LIU1116268/litewaverep.