32
Method : NFLD 分割
Dataset: Tajimi
Architecture: FCN-8s
Results: 最高 SE 98% , FP (5.42)
automated scheme for detection of a retinal nerve fiber layer defect(NFLD)
Method
Previous study
- multi-step detection : Gabor filtering , clustering and adaptive thresholding
- Problem :FP 多,method included too many rules
Deep convolutional neural network with fully convolutional layers
-
end-to-end : DCNN ( FCN-8s)
-
image
- Original color images ofabnormal cases,
- (b) original color images of both normal and abnormal cases,
- © ellipse-based polar transformed colorimages,
- (d) transformed G plane images,
- (e) transformed Gabor filtered color images,
- (f) transformed color halved images,
- (g) transformed color halved images with different data augmentation.
rotation and intensity transformation for all
-
add a softmax layer
Result
- 使用 normal and abnormal image 減少了 FP 但是也降低了 SE
- 雖然 綠色通道 對比度更高,但是用RGB圖具有更高的靈敏度
- DCNN 通用性比 previous study 更好
- FROC ,最高SE 98% , FP (5.42)
Discussion
- 研究了 不同輸入圖像 ,結果的不同,主要是從輸入數據這塊 做了一些對比實驗(個人感覺含金量不高)
- 研究了 FCN 在 眼底圖中自動檢測 NFLDs的應用