Abstract: To address the challenge of accurate joint segmentation of the Optic Cup (OC) and Optic Disc (OD) in early glaucoma diagnosis due to blurred boundaries and varied morphologies, a lightweight multi-scale feature enhancement network (LFM-Net) was proposed. This network employs an encoder-decoder architecture, aiming to improve segmentation accuracy by enhancing multi-scale feature representation and cross-layer feature fusion capabilities. Specifically, in the encoder stage, global contextual features were extracted layer by layer using depthwise separable convolutions and inverted bottleneck structures. A Multi-Scale Feature enhancement Aggregation (MSFA) module was introduced, utilizing its multi-branch convolutions and channel attention mechanisms to adaptively capture and aggregate global contextual and local detail features while maintaining low computational cost, thus addressing the significant differences in optic cup and optic disc sizes. In the decoder stage, a Convolutional Attention Feature Fusion Module (CAFM Fusion) was designed, combining 3D convolutional attention and pixel attention mechanisms to optimize feature transfer in skip connections, effectively suppressing background noise and sharpening edge responses, ultimately achieving efficient fusion of cross-layer features. Experimental results on three publicly available fundus image datasets—REFUGE, DRISHTI-GS, and RIM-ONE-r3—show that LFM-Net outperforms comparable methods such as U-Net, and TransUnet in key metrics including Dice coefficient, Intersection over Union (IoU), and accuracy. While maintaining lightweight design, LFM-Net accurately extracts OC and OD features, achieving high-precision segmentation. Furthermore, it demonstrates strong generalization ability across different datasets, providing effective technical support for computer-aided diagnosis of glaucoma.

Key words: glaucoma diagnosis, deep learning, optic cup and disc segmentation, multi-scale feature fusion, attention mechanism

摘要： 针对青光眼早期诊断中眼底图像视杯（OC）视盘（OD）因边界模糊、形态多变导致精准联合分割困难的挑战，
提出一种轻量化多尺度特征增强网络(LFM-Net)。该网络采用编码器-解码器架构，旨在通过增强多尺度特征表达与跨层特征融合能力以提升分割精度。具体实现上，编码器阶段利用深度可分离卷积和倒置瓶颈结构逐层提取全局上下文特征，并引入多尺度特征增强聚合（MSFA）模块，利用多分支卷积与通道注意力机制，在保持低计算成本的同时自适应捕获并聚合全局上下文与局部细节特征，以应对视杯视盘尺寸差异显著的问题；解码器阶段设计卷积注意力特征融合模块（CAFM Fusion），结合3D卷积注意力与像素注意力机制，优化跳跃连接中的特征传递，有效抑制背景噪声并锐化边缘响应,最终实现跨层特征的高效融合。在REFUGE、DRISHTI-GS和RIM-ONE-r3三个公开眼底图像数据集上的实验结果表明，LFM-Net在Dice系数、交并比（IoU）及准确度等关键指标上均优于U-Net、TransUnet等对比方法，在保证轻量化的同时能够准确提取OC和OD特征，实现高精度分割，并在跨数据集场景下展现出强泛化能力，为青光眼计算机辅助诊断提供了有效技术支持。

关键词: 青光眼诊断, 深度学习, 视杯视盘分割, 多尺度特征融合, 注意力机制