基於雙重注意力機制之視網膜血管分割深度學習網路

Abstract

眼底影像之血管分割可以用來協助眼睛病灶的觀察，以提早發現病灶並進行治療，例如黃斑部病變、糖尿病視網膜病變、青光眼等等。由於眼底影像的採集會經過各種不同的程序而導致影像有不同的品質變化，眼底影像血管分割的精確度會影響病灶的判斷，儘管現今已存在許多影像分割方法，但是具有病灶的眼底圖像血管分支變化多端，現存各種分割方法的精確度也依舊無法達到完美，本研究目的為提出改良式眼底影像的血管分割方法，針對各種視網膜圖像，進行精確血管分割，以協助醫師對眼疾病變的診斷，期能對眼疾醫療做出微薄的貢獻。準確的血管分割是一項具有挑戰性的任務，主要是因為眼底影像的對比度低以及血管形態結構的複雜性，傳統卷積會增加乘法的數量，同時執行卷積操作，導致與細長且對比度低的血管相關信息損失。為了解決現有方法在血管提取時低敏感度以及信息損失的問題，本研究提出結合兩種注意力模型EPA以及DLA的並行注意力U-Net以實現準確的血管分割，EPA聚焦於空間以及通道的特徵提取，而DLA則專注於多尺度的局部特徵以及邊緣檢測的特徵，再將並行所得特徵進行深度和淺層特徵融合。本研究在DRIVE數據集上進行實驗，以驗證模型性能，研究結果指出，採用並行運算的U-Net模型分割視網膜血管具有競爭性效能。Retinal vessel segmentation is a key step for the early diagnosis of fundus diseases, such as macular degeneration, diabetic retinopathy, and glaucoma. However, the accuracy of vessel segmentation is affected by variations in image quality due to different acquisition procedures. Deep learning-based retinal vessel segmentation has been shown to achieve feasible performance than traditional methods. However, most deep learning-based methods may still suffer from insufficiently capturing global and local features simultaneously from fundus images. This may degrade the segmentation performance, resulting in infeasible diagnosis for fundus diseases. To solve this problem, this thesis introduces the PADU-Net, a parallel attention-based dual U-Net architecture for retinal vessel segmentation. The key is to integrate two parallel U-Net modules, i.e., encoder-decoder architectures, equipped with local and global attention modules, respectively, used for extracting local and global features. Then the features are decoded and fused for generating the segmentation map for the input fundus image. The experiments conducted on the well-known dataset, DRIVE (digital retinal images for vessel extraction), has verified the performance of the proposed framework, outperforming the state-of-the-art methods.