Proceedings Article | 11 August 2023
KEYWORDS: Data modeling, Education and training, Optical coherence tomography, Performance modeling, Retinal diseases, Image classification, Deep learning, Eye models, Medical imaging, Machine learning
Optical Coherence Tomography (OCT) is a widely used medical imaging technique for diagnosing various eye conditions and for assessing the health of the retina and other structures in the eye. As the amount of data generated from OCT scans increases day by day, the need for automated support systems for medical staff becomes more pressing. In this study, we used a computer based diagnostic system for detecting and classifying three retinal diseases Choroidal Neovascularization (CNV), Drusen, Diabetic Macular Edema (DME)) in OCT B-scans. Our system applies Convolutional Neural Networks (EfficeintNet) and Transfer learning to make accurate predictions of the diseases or healthy eye. Transfer learning, a technique commonly used in deep learning, is implemented to train EfficientNet architectures using a publicly available dataset (OCT2017) of approx 84000 OCT B-scans. EfficientNet models were chosen for their known efficiency in terms of performance and their ability to classify effectively with fewer parameters and lower computational requirements in comparison to other Convolutional Neural Networks (CNN) models. We used EfficeintNet models that were pretrained on ImageNet dataset and fine-tuned the models on our dataset. In order to assess the effectiveness of transfer learning and fine-tuning, we evaluated their performance on an imbalanced multiclass classification task using metrics like F1 Score, Precision, Recall, Accuracy, and Confusion Matrices (CM). The results of our evaluations are presented in the form of CM for each class and model. We tried out various EfficientNet models (B0, B3, B5) as base models utilizing similar data resolution, volume and other hyperparameters, and discovered that larger EfficientNet models did not necessarily lead to better classification performance. We accounted for class imbalance in the data to make our method robust for real-world scenarios. The best result was found to be from lowest complexity model, EfficientNet B0. Our research found that the EfficientNet B0 model demonstrated exceptional performance with a macro average F1 Score of 99.8% and an Accuracy of 99.8%. Additionally, our results also revealed that the EfficientNets B0, B3, B5 models are particularly well-suited for multiclass classification based on highly imbalanced OCT2017 dataset. High classification scores are achieved due to several factors, such as data enhancement, resolution scaling, fine-tuning, and successful transfer learning using ImageNet weights. Based on our preliminary results our approach performs as well or outperforms other known approaches. Our goal is to provide assistance to medical staff in diagnositic process with the help of artificial intelligence (AI) algorithms. Improving the efficiency and accuracy of diagnosis is important in the field of medicine, and the use of AI algorithms such as the one proposed in this research has the potential to make a significant impact in this regard.
