A Dual Pooling Encoder U-Net for Retinal Vessel Segmentation
A Dual Pooling Encoder U-Net for Retinal Vessel Segmentation
- 주제(키워드) Dual Pooling Encoder , Medical Image Segmentation , Retinal Vessel Segmentation , U-Net , Weighted Average Mechanism
- 주제(DDC) 510
- 발행기관 아주대학교 일반대학원
- 지도교수 Dongwook Shin
- 발행년도 2025
- 학위수여년월 2025. 8
- 학위명 석사
- 학과 및 전공 일반대학원 수학과
- 실제URI http://www.dcollection.net/handler/ajou/000000035147
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Retinal vessel segmentation serves as a crucial image-based tool for the early diagnosis and prognosis of systemic and ocular diseases. Conventional U-Net models typically adopt Max Pooling for downsampling, which is advantageous for capturing vessel boundaries but limited in preserving the continuity of fine vascular structures. To address this limitation, we propose DPEU-Net, an extended U-Net architecture that introduces a Dual Pooling Encoder leveraging the complementary characteristics of Max Pooling and Average Pooling in parallel. The outputs are integrated through a Weighted Average Mechanism that aims to balance boundary recognition and structural continuity while suppressing false positives from background noise. Experimental results on the DRIVE and CHASE_DB1 datasets demonstrate that DPEU-Net outperforms the conventional U-Net using only Max Pooling in terms of accuracy, F1-score, AUC, and specificity. The segmented vessel maps exhibit clearer boundaries and improved continuity. Furthermore, when the proposed Dual Pooling Encoder and Weighted Average Mechanism were applied to Residual U-Net, Attention U-Net, and U-Net++, consistent improvements in key metrics were observed without major architectural modifications. These findings demonstrate the structural scalability and generalizability of the proposed method. Overall, DPEU-Net presents a practical and structurally flexible approach to enhancing retinal vessel segmentation and shows strong potential for adaptation to a wider range of medical image segmentation tasks.
more목차
1. Introduction 1
2. Methods 4
2.1 Dual Pooling Encoder 4
2.2 Weighted Average Mechanism 5
2.3 DPEU-Net Architecture 7
3. Experiments 9
3.1 Datasets and Preprocessing 9
3.2 Evaluation Metrics 11
3.3 Baseline Models 13
3.4 Experimental Setup 14
4. Results and Discussion 15
4.1 Comparison between U-Net and DPEU-Net 15
4.2 Comparison between U-Net Variants with and without DPEU-Net 19
5. Conclusions 21
References 23
국문요약 26
