PaperlayerView exact revision on arXiv

Source-linked AI summary

Transfer Learning for Automated OCTA Detection of Diabetic Retinopathy

David Le, Minhaj Alam, Cham Yao, Jennifer I. Lim, R. V. P. Chan, Devrim Toslak, Xincheng Yao

arXiv:1910.01796v1q-bio.QMeess.IV

TL;DR

Limited OCTA datasets make deep-learning classification of diabetic retinopathy challenging. The study uses transfer learning with VGG16 and integrates the retrained classifier into a GUI; the selected nine-layer model reports strong cross-validation performance, while the study identifies single-device data as a boundary requiring further validation.

  • Problem

    Limited OCTA datasets constrain evaluation of deep-learning classification for DR, despite OCTA's ability to show subtle retinal vascular changes.

    Source: INTRODUCTION · p. 3, DISCUSSION · pp. 13–14

  • Method

    The study transfers pretrained VGG16 weights, incrementally retrains upper layers for OCTA classification, and integrates the selected model into a GUI.

    Source: INTRODUCTION · p. 4, Classification Model Implementation · p. 6, DISCUSSION · p. 14

  • Results

    87.277% ACC, 83.764% SE, and 90.825% SP were achieved across control, NoDR, and DR on Dataset 1 with the upper nine layers retrained.

    Source: DISCUSSION · p. 14

  • Takeaways & Limitations

    The selected model and custom GUI demonstrated feasibility for exploring AI-based OCTA classification in a clinical environment.

    Source: DISCUSSION · p. 13

    The study used OCTA data from one device, and multi-device datasets are required for further validation.

    Source: DISCUSSION · p. 15

Abstract

from arXiv · show

Purpose: To test the feasibility of using deep learning for optical coherence tomography angiography (OCTA) detection of diabetic retinopathy (DR). Methods: A deep learning convolutional neural network (CNN) architecture VGG16 was employed for this study. A transfer learning process was implemented to re-train the CNN for robust OCTA classification. In order to demonstrate the feasibility of using this method for artificial intelligence (AI) screening of DR in clinical environments, the re-trained CNN was incorporated into a custom developed GUI platform which can be readily operated by ophthalmic personnel. Results: With last nine layers re-trained, CNN architecture achieved the best performance for automated OCTA classification. The overall accuracy of the re-trained classifier for differentiating healthy, NoDR, and NPDR was 87.27%, with 83.76% sensitivity and 90.82% specificity. The AUC metrics for binary classification of healthy, NoDR and DR were 0.97, 0.98 and 0.97, respectively. The GUI platform enabled easy validation of the method for AI screening of DR in a clinical environment. Conclusion: With a transfer leaning process to adopt the early layers for simple feature analysis and to re-train the upper layers for fine feature analysis, the CNN architecture VGG16 can be used for robust OCTA classification of healthy, NoDR, and NPDR eyes. Translational Relevance: OCTA can capture microvascular changes in early DR. A transfer learning process enables robust implementation of convolutional neural network (CNN) for automated OCTA classification of DR.

INTRODUCTION

The introduction motivates OCTA as a sensitive, noninvasive way to image retinal vasculature and frames limited OCTA data as a challenge for deep-learning classification. The study applies transfer learning to OCTA-based DR detection and develops a GUI to explore clinical use.

INTRODUCTION

  • OCTA visualizes retinal vasculature noninvasively and is more sensitive than fluorescein angiography for subtle vascular distortions associated with early eye conditions.The introduction also emphasizes early detection and treatment to prevent DR-related vision loss.Source: INTRODUCTION · pp. 2–3
  • Deep learning for OCTA classification is constrained by limited image availability, while CNN training for a specific task requires millions of images.The paper notes that this new imaging modality poses an obstacle to practical implementation.Source: INTRODUCTION · pp. 3–4
  • Transfer learning reuses weights from a pretrained CNN and retrains it for a specific task, with prior applications in fundus photography and OCT.The cited applications include artery-vein segmentation, glaucoma detection, diabetic macular thinning, CNV, DME, and AMD.Source: INTRODUCTION · p. 4
  • The study proposes transfer learning for OCTA-based DR classification and a custom GUI to assess its potential for clinical deployment.The introduction notes that the number of layers to retrain may depend on the dataset and task.Source: INTRODUCTION · p. 4

METHODS

The study reports adherence to the Declaration of Helsinki and approval by the University of Illinois at Chicago institutional review board.

METHODS

  • The study adhered to the Declaration of Helsinki and was approved by the University of Illinois at Chicago institutional review board.The approval statement names UIC's institutional review board.Source: METHODS · p. 5

Data Acquisition

OCTA images were acquired on an ANGIOVUE system, screened for significant artifacts, and collected from clinic patients and healthy volunteers under stated eligibility criteria.

Data Acquisition

  • OCTA data were acquired with a 6 x 6 mm2 field of view on an ANGIOVUE spectral-domain system, and images with significant motion or shadow artifacts were excluded.ReVue export and custom Python procedures were used for image processing.Source: Data Acquisition · p. 5

Patient Demographics

  • Patients were recruited from the UIC retina clinic, classified by two board-certified retina specialists using ETDRS staging, and eyes with other ocular disease or specified prior treatments were excluded.Healthy controls were volunteers who consented to OCT/OCTA imaging; demographic characteristics are summarized in Table 1.Source: Patient Demographics · pp. 5–6, Table 1 · p. 6

Classification Model Implementation

The study implements a VGG16 classifier initialized with ImageNet weights and trained with augmentation and input resizing.

Classification Model Implementation

  • VGG16 initialized with ImageNet weights was trained using stochastic gradient descent, early stopping, and random rotations, flips, and zoom augmentation.Training used batches of 12, a learning rate of 0.0001, and images resized from 304 x 304 to 224 x 224 pixels to meet VGG16 input specifications.Source: Classification Model Implementation · p. 6

Statistics

The study assessed classifier performance using AUC, sensitivity, specificity, accuracy, and misclassification error, with fivefold cross-validation. It selected a model using misclassification error and the one-standard-deviation rule.

Statistics

  • AUC, sensitivity, specificity, diagnostic accuracy, and misclassification error were used to assess model performance against manual classifications.ROC curves and AUC were determined using manual classifications as the reference standard.Source: Statistics · p. 7
  • Accuracy and misclassification error were calculated from true and false positive and negative counts.The passage gives formulas for accuracy and misclassification error using TP, TN, FP, and FN.Source: Statistics · p. 7
  • Fivefold cross-validation split the training dataset into five subsets, using four for training and one for validation.Source: Statistics · p. 7
  • The one-standard-deviation rule selected the model with the fewest retrained layers whose misclassification error was within one positive standard deviation of the best model.The passage relates this choice to the layer hierarchy: early layers extract simple features, while later layers extract complex features.Source: Statistics · p. 8

GUI/Software Integration

The selected model’s weights were exported and integrated into a Java/Python GUI, while training and evaluation used a separate Python environment.

GUI/Software Integration

RESULTS

The study included 131 OCTA images from control, NoDR, and DR patients, then augmented those images to create a larger dataset.

RESULTS

  • The study included 24 eyes from 17 NoDR patients, 75 eyes from 60 DR patients, and 32 eyes from 20 control subjects.Source: RESULTS · p. 8
  • Random flips, rotations, and zooms expanded the 131 raw OCTA images into 3,930 augmented images.Source: RESULTS · pp. 8–9
  • The raw OCTA images were designated Dataset 1, while the augmented images were designated Dataset 2.Source: RESULTS · p. 9

Model Selection

The nine-layer model was selected because its error met the one-standard-deviation criterion on both datasets, relative to the best-performing 15-layer model.

Model Selection

  • Models were evaluated on raw and augmented OCTA images to assess real-world performance and performance on a larger dataset.Source: Model Selection · p. 9
  • The VGG16 layers were sequentially retrained from the last layer toward the beginning, comparing one to sixteen layers.The figure counts convolutional and dense layers, excluding max-pooling and flattening operations because they have no optimizable parameters.Source: Model Selection · p. 9, Figure 2 · p. 9
  • Nine retrained layers met the one-standard-deviation criterion on both datasets and were selected, while the 15-layer model performed best.Figure 2 compares misclassification errors for models retrained with one to sixteen layers; nine layers was within one positive standard deviation of the 15-layer model.Source: Model Selection · pp. 9–10, Figure 2 · p. 9

Cross-validation Study

The selected model was evaluated with fivefold cross-validation on Dataset 1 and Dataset 2. The study reports category predictions, evaluation metrics, and ROC/AUC performance.

Cross-validation Study

Evaluation of Clinical Deployment

The best-performing model was integrated into a custom GUI to assess clinical use of OCTA-based DR classification. Ophthalmologists tested the platform, where classification takes under one minute per image.

Evaluation of Clinical Deployment

DISCUSSION

The authors report robust OCTA classification using a nine-layer-retrained VGG16 model and describe a custom GUI as a step toward clinical use. They note that limited, single-device OCTA data constrain validation.

DISCUSSION

  • The authors describe the custom GUI as a way to assess clinical deployment and a potential practical solution for DR mass-screening.They state that the platform may reduce the burden on experienced ophthalmologists.Source: DISCUSSION · p. 13
  • For this OCTA dataset, retraining nine upper VGG16 layers was the simplest model within one positive standard deviation of the fifteen-layer model’s misclassification error.The selection followed the one-standard-deviation rule and an incremental comparison of retraining from one to sixteen layers.Source: DISCUSSION · p. 14
  • Using the upper-nine-layers-retrained model, cross-validation averaged 87.277% ACC, 83.764% SE, and 90.825% SP across control, NoDR, and DR on Dataset 1; Dataset 2 achieved 90.789% ACC, 81.927% SE, and 90.894% SP.Overall AUC was 0.96 for Dataset 1 and 0.95 for Dataset 2.Source: DISCUSSION · p. 14
  • Current validation is constrained by limited OCTA datasets from one device; multi-device datasets are required for further validation.Source: DISCUSSION · p. 15

Related papers

Loading 1910.01796v1…