Purpose
The progression of age‐related macular degeneration (AMD) is critical to treatment decisions in clinical practice. The disease can be classified into four categories namely, drusen, inactive choroidal neovascularization (CNV), active CNV, and normal, according to severity based on optical coherence tomography (OCT) images. Interpreting numerous OCT images is still time‐consuming and labor‐intensive, especially for the detection of the CNV activity. To address this problem, we developed a deep learning (DL) system based on OCT images, with the assistance of an attention mechanism, to automatically diagnose AMD.
Methods
A public dataset (total 51,140 OCT images) and a private dataset (total 4951 OCT images) were utilized as a training dataset and a clinical validation dataset, respectively, to develop the DL model. A ResNet‐34 DL model, with convolutional block attention module (CBAM) block integrated into each unit, was pretrained on the public dataset first and then finetuned on our private dataset to automatically diagnose AMD and assist clinical decision‐making. GradCam, a visualization technique, was used to improve the interpretability of our model.
Results
The precision and recall of our model were, respectively, 84.3% and 87.3% for drusen, 81.2% and 80.0% for inactive CNV, 97.7% and 90.2% for active CNV, and 93.7% and 96.5% for normal. The area under the curve (AUC) corresponding to drusen, inactive CNV, active CNV, and normal for our model reached 0.9395, 0.9476, 0.9880, and 0.9925, respectively. The heatmaps indicated a high level of correspondence in the region of interest between our model and ophthalmologists on the diagnosis.
Conclusions
The implementation of finetuning and attention mechanism improve the performance of our model in a distinct dataset. Our model successfully assisted in the diagnosis of AMD and achieved a detection precision and recall equal to those of ophthalmologists. The results of our study could contribute to the precise diagnosis of and decision‐making regarding AMD.