The implementation of clinical-decision support algorithms for medical imaging faces challenges with reliability and interpretability. Here, we establish a diagnostic tool based on a deep-learning framework for the screening of patients with common treatable blinding retinal diseases. Our framework utilizes transfer learning, which trains a neural network with a fraction of the data of conventional approaches. Applying this approach to a dataset of optical coherence tomography images, we demonstrate performance comparable to that of human experts in classifying age-related macular degeneration and diabetic macular edema. We also provide a more transparent and interpretable diagnosis by highlighting the regions recognized by the neural network. We further demonstrate the general applicability of our AI system for diagnosis of pediatric pneumonia using chest X-ray images. This tool may ultimately aid in expediting the diagnosis and referral of these treatable conditions, thereby facilitating earlier treatment, resulting in improved clinical outcomes. VIDEO ABSTRACT.

[This corrects the article DOI: 10.1038/sigtrans.2016.16.].

In order to improve the accuracy of convolutional neural networks (CNN) in imbalanced dataset classification, a novel hierarchical CNN-LSVM is proposed. Considering the imbalance in the number and spatial distribution of wheat leaf disease images, the improved local support vector machine (LSVM) replaces Softmax as the classifier of the model, and meanwhile a cost sensitive matrix is designed to assign the value for penalty factors in the optimized objective function of LSVM. It effectively improves the sensitivity of misclassification caused by imbalanced data. To verify the validity and practicability of CNN-LSVM, 6028 wheat leaf disease images containing 8 species are collected from planting bases in Shandong Agricultural University. Then the imbalanced and balanced standard image sets are generated by data augmentation and Borderline-Synthetic Minority Oversampling (Borderline-SMOTE). They have 36168 and 46176 images, respectively. The experimental results show that the average identification accuracies of the CNN-LSVM obtained on imbalanced and balanced standard datasets are 90.32 % and 93.68 %, respectively. And it starts to converge when the iteration times are close to 13000. CNN-LSVM has higher classification accuracy and lower iteration times, compared with CNN-Softmax, CNN-SVM, LSVM and support vector machine (SVM).

Age-related macular degeneration (AMD) is characterized by complex interactions between genetic and environmental factors. Here we genotyped the selected 25 single-nucleotide polymorphisms (SNPs) in 983 cases with advanced AMD and 271 cases with intermediate AMD and build an AMD life-risk score model for assessment of progression from intermediate to advanced AMD. We analyzed the performance of the prediction model for geographic atrophy progressors or choroidal neovascularization progressors versus non-progressors based on the 25 SNPs plus body mass index and smoking status. Our results suggest that a class prediction algorithm can be used for the risk assessment of progression from intermediate to late AMD stages. The algorithm could also be potentially applied for therapeutic response, and toward personalized care and precision medicine.

No abstract