A Computationally Efficient Deeplabv3+ Framework With A Mobilenetv2 Backbone For Accurate Retinal Vessel Segmentation In Fundus Images

Abstract

Segmentation of retinal vessels is an essential task in the computer-aided diagnosis of several ophthalmic diseases like DR, glaucoma and hypertensive retinopathy. The current deep learning models have several drawbacks, such as high computational complexity, large memory requirements, and are not suitable for real-time deployment in clinical settings on resource-constrained devices. In this work, a computationally efficient DeepLabV3+ network with MobileNetV2 backbone is proposed to perform automatic retinal vessel segmentation in fundus images overcoming these issues. The proposed framework is based on a lightweight MobileNetV2 encoder, Atrous Spatial Pyramid Pooling (ASPP) and an optimized decoder that can effectively capture the context features while minimizing the computational burden. The FIVES (Fundus Image Vessel Segmentation) dataset, was used for developing and testing the model. Data was resized, transformed into the RGB color space, and enhanced for contrast, with a large amount of data augmentation performed, and the model optimized with the Adam optimizer and a Weighted BCE–Dice loss function, tackling foreground–background class imbalance. Experiments were run on U-Net and DeepLabV3+ Backbone. The proposed framework had 3.205 million trainable parameters, which is 20% fewer than the 7.703 million and 59.341 million parameters for U-Net and DeepLabV3+ (ResNet101), respectively, and only 12.82 MB memory consumption for parameters, putting it 30 times below the 400 MB of U-Net and 3,553 MB of DeepLabV3+ (ResNet101), respectively, and 4.03 GB multi-add operations per inference, which is 10 times lower than the 42.4 GB for U-Net and 355.3 GB for DeepLabV3+ (ResNet101), respectively. The proposed framework is lightweight and computationally efficient solution for retinal vessel segmentation which provides a practical basis for accurate, real-time, edge-based ophthalmic image analysis with enhanced deployment practicability.Segmentation of retinal vessels is an essential task in the computer-aided diagnosis of several ophthalmic diseases like DR, glaucoma and hypertensive retinopathy. The current deep learning models have several drawbacks, such as high computational complexity, large memory requirements, and are not suitable for real-time deployment in clinical settings on resource-constrained devices. In this work, a computationally efficient DeepLabV3+ network with MobileNetV2 backbone is proposed to perform automatic retinal vessel segmentation in fundus images overcoming these issues. The proposed framework is based on a lightweight MobileNetV2 encoder, Atrous Spatial Pyramid Pooling (ASPP) and an optimized decoder that can effectively capture the context features while minimizing the computational burden. The FIVES (Fundus Image Vessel Segmentation) dataset, was used for developing and testing the model. Data was resized, transformed into the RGB color space, and enhanced for contrast, with a large amount of data augmentation performed, and the model optimized with the Adam optimizer and a Weighted BCE–Dice loss function, tackling foreground–background class imbalance. Experiments were run on U-Net and DeepLabV3+ Backbone. The proposed framework had 3.205 million trainable parameters, which is 20% fewer than the 7.703 million and 59.341 million parameters for U-Net and DeepLabV3+ (ResNet101), respectively, and only 12.82 MB memory consumption for parameters, putting it 30 times below the 400 MB of U-Net and 3,553 MB of DeepLabV3+ (ResNet101), respectively, and 4.03 GB multi-add operations per inference, which is 10 times lower than the 42.4 GB for U-Net and 355.3 GB for DeepLabV3+ (ResNet101), respectively. The proposed framework is lightweight and computationally efficient solution for retinal vessel segmentation which provides a practical basis for accurate, real-time, edge-based ophthalmic image analysis with enhanced deployment practicability.