- 學位論文
改良U-Net用於醫學影像分割輔助診斷乾癬面積
Improved U-Net for Medical Image Segmentation to Assist Diagnosis of Psoriasis Area
摘要
U-Net架構常用於醫學影像分割，主要好處是簡單、容易構建且適用於不同尺度的影像。醫學影像內容比較固定且單一，所有特徵都是重要的，不適合刪減資訊，而傳統U-Net架構的編碼區塊過於簡單，在乾癬的特徵學習較沒有效率且分割能力不佳。為了改善語意分割能力，本論文改進編碼區塊與解碼區塊，當中也增加注意力模塊以加強資訊的結合。使用修改的架構分割的病例影像，在乾癬分割的並交比（Intersection over Union, IOU）提升近10%，可用於計算身體各部位乾癬與皮膚的像素點占比，再依據體表面積(Body Surface Area, BSA)計算嚴重度。
並列摘要
The U-Net architecture is often used for medical image segmentation. The main advantage is that it is simple, easy to construct and suitable for images of different scales. The medical image content is relatively fixed and single, all features are important, and it is not suitable for deleting information. The coding block of the traditional U-Net architecture is too simple, and the feature learning in psoriasis is less efficient and the segmentation ability is poor. In order to improve the semantic segmentation ability, this paper improves the coding block and the decoding block, and also adds a attention module to strengthen the combination of information. Case images segmented using the modified architecture have increased the cross-over ratio of psoriasis segmentation by nearly 10%, which can be used to calculate the proportion of psoriasis and skin pixels in various parts of the body, and then calculate the severity based on the Body Surface Area.