Gray’s Anatomy for Segment Anything Model: Optimizing Grayscale Medical Images for Fast and Lightweight Segmentation
Keywords: Medical imaging, Segmentation, Lightweight
Abstract: Advancements in medical image segmentation are critical for enhancing diagnostic accuracy in clinical settings, particularly when operating on edge devices like CPU-only laptops. In this context, we have developed a medical image segmentation model that is specifically designed for efficient deployment on such devices. Our approach leverages the EfficientViT-SAM architecture integrated with dynamic quantization to optimize both accuracy and computational efficiency. The model has been trained on a diverse dataset that includes over one million image-mask pairs from 10 different medical imaging modalities along with additional data for underrepresented anatomies. Performance evaluations show that our model achieves a dice score of 88.54\% and a normalized surface dice of 98.28\%, showing improvements of 4.37\% and 2.85\%, respectively, over the baseline model. The implementation of dynamic quantization not only preserves accuracy but also boosts inference speeds, making the model exceptionally viable for real-time clinical applications. This study affirms the potential of advanced segmentation technologies to operate effectively on non-specialized hardware, thereby expanding the accessibility of high-quality medical imaging analysis in environments constrained by resources. With its robust performance across various imaging scenarios and enhanced processing efficiency, the model promises substantial improvements in clinical workflows and patient outcomes. The code is available at https://github.com/Ninebell/GraysAnatomySAM.
Submission Number: 12
Loading