Abstract
Purpose :
To apply transfer learning from the domain of diabetic retinopathy (DR) to aid in the detection of AMD
Methods :
The proposed model was based on the ResNet-RS architecture. Pre-training aimed at DR diagnosis was conducted using the Kaggle database (EyePACS) released for DR research, consisting of more than 80,000 fundus images but after eliminating those of por quality 52,973 images were finally used and divided into a training set of 52,073 images (13,099 with DR and 38,974 with no DR) and a validation set of 900 images (200 with DR and 700 with no DR).
A preprocessing stage for input normalization was introduced and after a deep CNN backbone was linked to a set of fully-connected layers.
Then, fine-tunning was performed using other public data set, the ADAM dataset that contains 1,200 fundus images provided by the Zhongshan Ophthalmic Center of the Sun Yat-sen University (China). Only a small dataset of 400 images from ADAM was considered in this work.
We carried out 3 experiments with different number of images , N = 50, 100 and 400 images, respectively, as the training set of the fine-tunning stage.
Results :
As the main result, our method showed a much faster convergence than the corresponding models pre- trained on ImageNet. When analyzing the experiments with N=50, we could observe how our proposal surpasses the traditional approach in every metric except for SP. In particular, the value of AUC=0.88 is notably higher than the AUC=0.78. As for the experiments with N=100, the conclusion seems to be similar, but the computed metrics are closer among them.
When N=400, however, the obtained results are on the same line. That is, our approach does not imply better performance for this case. These comparisons show that the proposed approach is especially useful when scarce data in the destination domain is available.
Conclusions :
The present work revealed that an appropriate source domain when applying transfer learning helps improve the model effectiveness and efficiency. In contrast to ImageNet pre-training, the use of a pre-trained network aimed at DR diagnosis allowed us to achieve a much faster convergence for AMD detection. Additionally, the proposed approach showed a higher performance when a reduced subset of images was available for fine-tunning purposes.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.