Abstract
Purpose :
While prognosis and risk of progression is crucial in developing precise therapeutic strategy in treatment-naive neovascular age-related macular degeneration (nAMD) patients, limited predictive tools are available. We proposed a novel deep convolutional neural network (CNN) that enables feature extraction through image and non-image data integration to seize imperative information and achieve highly accurate outcome prediction.
Methods :
A total of 698 evaluable nAMD patients who received intravitreal injection of anti–vascular endothelial growth factor (anti-VEGF) were analyzed. The Heterogeneous Data Fusion Net (HDF-Net) was designed to predict VA outcome in 12th months after anti-VEGF treatment. A set of baseline optical coherence tomography (OCT) image and non-image demographic features were employed as input data and the 12th-month VA as target data to train, validate, and test the HDF-Net. Accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) were evaluated.
Results :
Of all evaluable patients,163 patients had at least 2-line VA improvement, and 535 failed to achieve VA improvement (≥ 2 line).This newly designed HDF-Net was trained to perform 12-month VA forecast and demonstrated an AUC of 0.989 (95% CI, 0.970-0.999), accuracy of 0.936 (95% confidence interval [CI], 0.889-0.964), sensitivity of 0.933 (95% CI, 0.841-0.974), and specificity of 0.938 (95% CI, 0.877-0.969).The HDF-Net demonstrated superior performance to the classic AlexNet (AUC of 0.936 (95% CI, 0.894-0.978), accuracy of 0.895 (95% CI, 0.841-0.933), sensitivity of 0.824 (95% CI, 0.716-0.896), and specificity of 0.942 (95% CI, 0.880-0.973)). In the attention heatmap, focused locations that contributed most to decision making by HDF-Net showed validity and corresponded well to clinically relevant features within OCT images.
Conclusions :
By simulating the real-world clinical decision process with mixed baseline information from OCT images and non-image data, HDF-Net demonstrated promising performance in predicting individualized treatment outcome. The results highlight the potential of deep learning to simultaneously process a broad range of clinical data to weigh and leverage the complete information of the patient. This novel approach is an important step toward personalized therapeutic strategy for nAMD.
This is a 2021 ARVO Annual Meeting abstract.