FunduSAM: A Specialized Deep Learning Model for Enhanced Optic Disc and Cup Segmentation in Fundus Images

TL;DR

FunduSAM addresses the gap in applying foundation models to fundus OD/OC segmentation by introducing PEFT adapters in the ViT blocks of SAM, complemented by CBAM and polar preprocessing. A joint loss $L=\omega_1 L_{disk} + \omega_2 L_{cup} + \omega_3 L_{contain}$ enforces structural priors that OC lies within OD, improving consistency between the two segments. Evaluations on the REFUGE dataset show FunduSAM achieving state-of-the-art Dice and IOU for both OD and OC, outperforming ResUNet, nnUNet, TransUNet, Swin-UNetr, and MedSAM. The results demonstrate that a targeted, parameter-efficient adaptation of a foundation model, combined with domain-specific preprocessing and loss design, can yield robust medical image segmentation with practical clinical impact.

Abstract

The Segment Anything Model (SAM) has gained popularity as a versatile image segmentation method, thanks to its strong generalization capabilities across various domains. However, when applied to optic disc (OD) and optic cup (OC) segmentation tasks, SAM encounters challenges due to the complex structures, low contrast, and blurred boundaries typical of fundus images, leading to suboptimal performance. To overcome these challenges, we introduce a novel model, FunduSAM, which incorporates several Adapters into SAM to create a deep network specifically designed for OD and OC segmentation. The FunduSAM utilizes Adapter into each transformer block after encoder for parameter fine-tuning (PEFT). It enhances SAM's feature extraction capabilities by designing a Convolutional Block Attention Module (CBAM), addressing issues related to blurred boundaries and low contrast. Given the unique requirements of OD and OC segmentation, polar transformation is used to convert the original fundus OD images into a format better suited for training and evaluating FunduSAM. A joint loss is used to achieve structure preservation between the OD and OC, while accurate segmentation. Extensive experiments on the REFUGE dataset, comprising 1,200 fundus images, demonstrate the superior performance of FunduSAM compared to five mainstream approaches.

Figures (4)

• Figure 1: The pipeline of our method for optic disc and cup segmentation, the input fundus image is first transformed in polar coordinates, followed by learning and prediction of the image using our proposed FunduSAM, and finally the output is transformed in inverse polar coordinates to obtain the result
• Figure 2: Overview of the proposed FunduSAM consists of Adapter layers, Spatial Attention Module and Channel Attention Module in CBAM
• Figure 3: Visualization of polar transformation, which simplifies the boundary information and balances the proportion of OD and OC.
• Figure 4: Visualization of the results of different methods for fundus optic disc and cup segmentation