[Lesion detection in optical coherence tomography based on lightweight convolutional neural networks]

Objective: To achieve automated lesion detection in optical coherence tomography (OCT) images based on a lightweight convolutional neural network architecture. Methods: This retrospective study employed deep learning to construct a lightweight lesion identification model using OCT images from Sichuan Provincial People's Hospital alongside datasets including OCT-C8, OCT-2017, and HD-OCT of MH. Model performance evaluation comprised two stages: firstly, testing on four independent external validation sets to assess the model's generalisability and accuracy; secondly, employing a confidence interval overlap comparison method to evaluate the classification performance of ophthalmologists at different levels of clinical experience (two per level) against the model, thereby determining the clinical experience level corresponding to the model's medical proficiency. Results: The model achieved an accuracy of 93.41% (47 054/50 374) on the validation set, with an F1 score of 88.18% (44 419/50 374) and a recall rate of 86.42% (43 533/50 374). The area under the receiver operating characteristic curve was 99.31% (50 026/50 374), with precision at 91.48% (46 082/50 374) and specificity at 99.1% (49 921/50 374). For nine categories of OCT images, namely neovascularisation, posterior vitreous detachment, epiretinal membrane, macular hole, macular schisis, subretinal fluid, normal image, vitreomacular traction, and obscured image, the precision rates were 97.48% (16 873/17 310), 90.99% (3 039/3 340), 97.62% (1 325/1 358), 97.91% (2 185/2 232), 79.23% (4 961/6 262), 81.47% (1 594/1 957), 90.55% (17 744/19 596), 88.08% (237/269), and 99.95% (19 520/19 530), respectively. The model achieved an average accuracy of 89.9% on the external validation set, whereas junior physicians demonstrated an accuracy of 74% (95%CI: 61.84%~86.16%) in interpreting OCT images, and that of mid-level physicians was 88% (95%CI: 78.99%~97.01%), indicating the model's performance approached that of mid-level physicians. Conclusions: This study marks the first successful implementation of automated recognition across nine categories of OCT images. Its clinical performance has preliminarily attained the level of a mid-career physician, and it can be deployed locally within healthcare settings to enhance diagnostic efficiency and accuracy.