TY - GEN
T1 - DeepAMD
T2 - 14th Asian Conference on Computer Vision, ACCV 2018
AU - Liu, Huiying
AU - Wong, Damon W.K.
AU - Fu, Huazhu
AU - Xu, Yanwu
AU - Liu, Jiang
N1 - Publisher Copyright:
© Springer Nature Switzerland AG 2019.
PY - 2019
Y1 - 2019
N2 - Automatic screening of Age-related Macular Degeneration (AMD) is important for both patients and ophthalmologists. In this paper, we focus on the task of AMD detection at the very early stage from fundus images. The difficulty of this task is that at the very early stage, the signs, e.g., drusen, are too tiny and subtle to be detected by most of the current methods. To address this issue, we apply deep learning in a multiple instance learning framework to catch these subtle features to detect AMD at the very early stage. The deep networks is able to learn a discriminative representation of the subtle signs of AMD. The multiple instance learning framework helps in two ways. First, It is able to choose the location where AMD happens because it works on image patches instead of the whole image. Second, It works on the image of high resolution instead of down sampling the image which may lead to invisibility of the tiny drusen. The experiments are carried out on a dataset consists of 3596 AMD and 1129 normal fundus images. The final average AUC is 0.79, compared with 0.74 of the same neural network but without multiple instance learning.
AB - Automatic screening of Age-related Macular Degeneration (AMD) is important for both patients and ophthalmologists. In this paper, we focus on the task of AMD detection at the very early stage from fundus images. The difficulty of this task is that at the very early stage, the signs, e.g., drusen, are too tiny and subtle to be detected by most of the current methods. To address this issue, we apply deep learning in a multiple instance learning framework to catch these subtle features to detect AMD at the very early stage. The deep networks is able to learn a discriminative representation of the subtle signs of AMD. The multiple instance learning framework helps in two ways. First, It is able to choose the location where AMD happens because it works on image patches instead of the whole image. Second, It works on the image of high resolution instead of down sampling the image which may lead to invisibility of the tiny drusen. The experiments are carried out on a dataset consists of 3596 AMD and 1129 normal fundus images. The final average AUC is 0.79, compared with 0.74 of the same neural network but without multiple instance learning.
UR - http://www.scopus.com/inward/record.url?scp=85066783962&partnerID=8YFLogxK
U2 - 10.1007/978-3-030-20873-8_40
DO - 10.1007/978-3-030-20873-8_40
M3 - Conference contribution
AN - SCOPUS:85066783962
SN - 9783030208721
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 625
EP - 640
BT - Computer Vision – ACCV 2018 - 14th Asian Conference on Computer Vision, Revised Selected Papers
A2 - Schindler, Konrad
A2 - Jawahar, C.V.
A2 - Li, Hongdong
A2 - Mori, Greg
PB - Springer Verlag
Y2 - 2 December 2018 through 6 December 2018
ER -