Core Concepts
H-SAM introduces a prompt-free adaptation of SAM for efficient fine-tuning of medical images through a hierarchical decoding process, showcasing superior performance without relying on unlabeled data.
Abstract
Abstract:
SAM's versatile segmentation abilities and prompt-based interface.
Challenges in applying SAM to medical imaging.
Introduction of H-SAM for efficient fine-tuning of medical images.
Introduction:
Importance of accurate medical image segmentation.
SAM's limitations in medical image segmentation.
Introduction of H-SAM for enhanced adaptation in medical imaging.
Methodology:
Overview of H-SAM's hierarchical decoding process.
Details of LoRA-adapted image encoder and mask decoder.
Training loss and deep supervision techniques.
Experiments:
Evaluation on Synapse, LA, and PROMISE12 datasets.
Comparison with state-of-the-art models in both few-shot and fully-supervised settings.
Ablation Study:
Effectiveness of Learnable Mask Cross Attention, CMAttn, and Hierarchical Pixel Decoder.
Efficiency Analysis:
Comparison of total parameters and performance with other SAM variants.
Qualitative Results:
Visual comparison of H-SAM with other SAM variants in medical image segmentation.
Stats
H-SAM demonstrates a 4.78% improvement in average Dice compared to existing prompt-free SAM variants for multi-organ segmentation using only 10% of 2D slices.
H-SAM achieves a Mean Dice of 89.22% on the LA dataset using 4 labeled scans for training.
H-SAM shows a significant improvement in Dice coefficients on the PROMISE12 dataset, achieving 87.27% with only 3 labeled cases for training.
Quotes
"H-SAM surpasses existing prompt-free SAM variants for multi-organ segmentation with limited samples."
"H-SAM demonstrates superior performance without relying on any unlabeled data."