Deep learning-based single-shot prediction of differential effects of anti-VEGF treatment in patients with diabetic macular edema

Abstract

Anti-vascular endothelial growth factor (VEGF) agents are widely regarded as the first line of therapy for diabetic macular edema (DME) but are not universally effective. An automatic method that can predict whether a patient is likely to respond to anti-VEGF therapy can avoid unnecessary trial and error treatment strategies and promote the selection of more effective first-line therapies. The objective of this study is to automatically predict the efficacy of anti-VEGF treatment of DME in individual patients based on optical coherence tomography (OCT) images. We performed a retrospective study of 127 subjects treated for DME with three consecutive injections of anti-VEGF agents. Patients' retinas were imaged using spectral-domain OCT (SD-OCT) before and after anti-VEGF therapy, and the total retinal thicknesses before and after treatment were extracted from OCT B-scans. A novel deep convolutional neural network was designed and evaluated using pre-treatment OCT scans as input and differential retinal thickness as output, with 5-fold cross-validation. The group of patients responsive to anti-VEGF treatment was defined as those with at least a 10% reduction in retinal thickness following treatment. The predictive performance of the system was evaluated by calculating the precision, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). The algorithm achieved an average AUC of 0.866 in discriminating responsive from non-responsive patients, with an average precision, sensitivity, and specificity of 85.5%, 80.1%, and 85.0%, respectively. Classification precision was significantly higher when differentiating between very responsive and very unresponsive patients. The proposed automatic algorithm accurately predicts the response to anti-VEGF treatment in DME patients based on OCT images. This pilot study is a critical step toward using non-invasive imaging and automated analysis to select the most effective therapy for a patient's specific disease condition.

Fig. 1.

Example foveal SD-OCT images from pre-treatment (row 1) and post-treatment (row 2) acquisition sets. Only the patients in the second and third columns showed signs of response to treatment.

Fig. 2.

Histogram of retinal thickness changes in pre-treatment OCT B-scans. The horizontal axis indicates the central thickness difference between post-treatment and baseline screenings. (Left) Differential thickness (µm). (Right) Percentage change in differential thickness.

Fig. 3.

Overview of the CADNet predictive framework with m = 6 attention blocks. The SE-Unit is demonstrated in detail. Values inside the bracket indicate the kernel size and the number of feature maps according to the block number, respectively. RetiUnet is a developed and pre-trained UNet model used as a non-trainable layer of CADNet for total retina segmentation. The sub-sampling factor and squeeze ratio of the pooling layers and SE-Units were 2 and 8, respectively. The symbols ⊗ and ⊕ indicate element-wise multiplication and summation operations, respectively. (GAP: global average pooling layer; FC: fully connected; ReLU: rectified linear units)

Fig. 4.

Plot showing cross-validated precision performance against the epoch for the CADNet model. To avoid overfitting, we terminated the training process at the 50^th epoch, at which point the validation precision shows lower performance. Due to our limited database and the wide range of DME manifestations on OCT in this prediction problem, our model is prone to overfitting.

Fig. 5.

Performance of the CADNet + RFE.EN + GNB framework. (Left column) Results at the ROI level. (Right column) Results at the patient level. (Top row) ROC curves. (Bottom row) Confusion matrices.