Adapting Segment Anything for Retinal OCT Fluid Segmentation

Recent advancements in deep learning have highlighted the potential of segmentation models like the Segment Anything Model (SAM) to enhance segmentation performance in natural images. SAM, a versatile deep learning framework, excels in natural image contexts due to its ability to adapt flexibly without specific pre-training, utilizing user-provided prompts to segment a wide variety of objects. However, the application of SAM’s zero-shot capabilities to medical images presents challenges, primarily due to the significant differences between natural and medical imaging characteristics. Fine-tuning foundational models for medical imaging adaptation has been the standard approach. However, this process is time-consuming and computationally expensive, especially due to the shortage of good quality of annotated medical images and variability of medical images. To mitigate these challenges, adapters are lightweight modules integrated into foundational models, designed to learn domain-specific details. Existing adapters often fall short of capturing the intricate, data-specific features due to their simplistic design. To address these limitations, we propose a novel adapter that effectively learns and captures features unique to the medical domain, particularly for segmenting retinal fluids from Optical Coherence Tomography (OCT) data. Our adapter employs dilated convolutions combined with a shared parameter strategy to enhance multi-scale feature extraction, significantly improving the feature extraction pipeline by capturing details at varying scales. Notably, integrating our adapter into the SAM's encoder results in a 30% reduction in fine-tunable parameters compared to the SAM's encoder, enabling more efficient fine-tuning. Additionally, our approach demonstrates robustness across different prompting locations, leading to consistent improvements in segmentation metrics.