An Efficient ConvMixer with Mini-ASPP for Retinal Blood Vessel Segmentation

Abstract

Retinal blood vessel segmentation remains crucial for faster clinical decision-making regarding a patient's disease, enabling faster medical treatment. However, the complexity of tiny blood vessels makes segmentation challenging. This study employs a novel method using ConvMixer, which operates on the principle of an encoder and decoder, to process retinal blood vessel images. ConvMixer utilizes a depthwise process that focuses on mixing spatial locations and a pointwise process that focuses on mixing channel-wise image features. Furthermore, to enhance feature extraction and explore multiscale image processing in retinal blood vessels, we introduce Mini-ASPP as the bottleneck to capture multiscale details of blood vessels, thereby improving segmentation performance. To test the effectiveness of our architecture, we used the public retinal blood vessel dataset with a small subset of images to assess our network's segmentation efficiency. The results were a Dice coefficient of 0.8125 on the training data and 0.8065 on the testing data, along with relatively high recall and AUC-ROC values, indicating promising model performance. These results demonstrate that ConvMixer and Mini-ASPP can achieve competitive segmentation performance and are computationally efficient for the segmentation process. These results will help improve clinical decision-making in the future.