Novel Extraction of Discriminative Fine-Grained Feature to Improve Retinal Vessel Segmentation

TL;DR

The paper tackles pixel-wise retinal vessel segmentation by identifying underutilization of encoder fine-grained features in existing models. It introduces AttUKAN, a U-shaped network that combines Attention Gates with Kolmogorov-Arnold Network blocks, and augments it with LPCL to enforce discriminative feature-level representations via pixel-wise contrastive supervision. Across DRIVE, STARE, CHASE_DB1, HRF, and a private dataset, AttUKAN delivers state-of-the-art results and ablation analyses confirm the effectiveness of both Attention Gates and LPCL, including LPCL’s applicability to other networks. The proposed approach improves segmentation accuracy, especially for tiny vessels, and provides a reusable framework with practical implications for automated retinal analysis.

Abstract

Retinal vessel segmentation is a vital early detection method for several severe ocular diseases. Despite significant progress in retinal vessel segmentation with the advancement of Neural Networks, there are still challenges to overcome. Specifically, retinal vessel segmentation aims to predict the class label for every pixel within a fundus image, with a primary focus on intra-image discrimination, making it vital for models to extract more discriminative features. Nevertheless, existing methods primarily focus on minimizing the difference between the output from the decoder and the label, but ignore fully using feature-level fine-grained representations from the encoder. To address these issues, we propose a novel Attention U-shaped Kolmogorov-Arnold Network named AttUKAN along with a novel Label-guided Pixel-wise Contrastive Loss for retinal vessel segmentation. Specifically, we implement Attention Gates into Kolmogorov-Arnold Networks to enhance model sensitivity by suppressing irrelevant feature activations and model interpretability by non-linear modeling of KAN blocks. Additionally, we also design a novel Label-guided Pixel-wise Contrastive Loss to supervise our proposed AttUKAN to extract more discriminative features by distinguishing between foreground vessel-pixel pairs and background pairs. Experiments are conducted across four public datasets including DRIVE, STARE, CHASE_DB1, HRF and our private dataset. AttUKAN achieves F1 scores of 82.50%, 81.14%, 81.34%, 80.21% and 80.09%, along with MIoU scores of 70.24%, 68.64%, 68.59%, 67.21% and 66.94% in the above datasets, which are the highest compared to 11 networks for retinal vessel segmentation. Quantitative and qualitative results show that our AttUKAN achieves state-of-the-art performance and outperforms existing retinal vessel segmentation methods. Our code will be available at https://github.com/stevezs315/AttUKAN.

Figures (9)

• Figure 1: (a) A fundus image from IDRiD dataset illustrating important biomarkers and lesions. (b) An illustration of Diabetic Retinopathy fundus image. (c) An illustration of Hypertensive Retinopathy fundus image.
• Figure 2: Overview of AttUKAN pipeline. the input retinal vessel image is processed through the encoder, consisting of three convolutional blocks and two tokenized KAN blocks and the symmetric decoder. Attention Gates are incorporated into skip connection and the model is optimized with a hybrid baseline loss along with our proposed LPCL loss.
• Figure 3: Illustrations of each module in AttUKAN. (a) Convolution Block aims to extract features by leveraging convolutional operations. (b) Tokenized KAN Block implements KAN layers to augment the model's capacity for non-linear modeling. (c) Attention Gate aims to selectively filter and enhance feature-level fine-grained representations.
• Figure 4: Label-guided Pixel-wise Contrastive Loss. Red circles denote foreground vessel vectors, while blue triangles represent background vectors. Contrastive Learning Loss is employed to pull similar representations closer together and push dissimilar representations apart to supervise our AttUKAN learning more discriminative feature-level representation.
• Figure 5: Visualization of tiny vessel segmentation results on the DRIVE dataset. Green pixels represent true positive predictions; yellow pixels indicate false positive predictions; and red pixels denote false negatives predictions. Our proposed AttUKAN can get a more accurate segmentation result with more green regions and higher MIoU.