The classification of stages of epiretinal membrane using convolutional neural network on optical coherence tomography image

• Published in: Methods (San Diego, Calif.) Vol. 214; pp. 28 - 34
• Main Authors: Hung, Che-Lun, Lin, Keng-Hung, Lee, Yu-Kai, Mrozek, Dariusz, Tsai, Yin-Te, Lin, Chun-Hsien
• Format: Journal Article
• Language: English
• Published: United States Elsevier Inc 01.06.2023
• Subjects: Algorithms; Classification; Combintorial Fusion; Convolutional Neural Network; Epiretinal Membrane; Epiretinal Membrane - diagnostic imaging; Humans; Neural Networks, Computer; Optical Coherence Tomography; Retina; Tomography, Optical Coherence - methods; Convolutional Neural Network; ERM; OCT; CNN; ROC; DL; EIFL; ILM; AUC; MLP; SD-OCT; Optical Coherence Tomography; Classification; Combintorial Fusion; Epiretinal Membrane; CFA; PVD
• ISSN: 1046-2023; 1095-9130
• DOI: 10.1016/j.ymeth.2023.04.006

•Stages of Epiretinal Membrane can be classified with deep learning methods.•Fusing different convolutional neural networks increases classification accuracy.•Deep learning can support ophthalmologists in making diagnosing decisions.•Explainable artificial intelligence is necessary for ophthalmologists to verify the diagnosing features. The gold standard for diagnosing epiretinal membranes is to observe the surface of the internal limiting membrane on optical coherence tomography images. The stages of the epiretinal membrane are used to decide the condition of the health of the membrane. The stages are not detected because some of them are similar. To accurately classify the stages, a deep-learning technology can be used to improve the classification accuracy. Methods: A combinatorial fusion with multiple convolutional neural networks (CNN) algorithms are proposed to enhance the accuracy of a single image classification model. The proposed method was trained using a dataset of 1947 optical coherence tomography images diagnosed with the epiretinal membrane at the Taichung Veterans General Hospital in Taiwan. The images consisted of 4 stages; stages 1, 2, 3, and 4. Results: The overall accuracy of the classification was 84%. The combination of five and six CNN models achieves the highest testing accuracy (85%) among other combinations, respectively. Any combination with a different number of CNN models outperforms any single CNN algorithm working alone. Meanwhile, the accuracy of the proposed method is better than ophthalmologists with years of clinical experience. Conclusions: We have developed an efficient epiretinal membrane classification method by using combinatorial fusion with CNN models on optical coherence tomography images. The proposed method can be used for screening purposes to facilitate ophthalmologists making the correct diagnoses in general medical practice.