Research on deep learning-based lesion identification in optical coherence tomography

Abstract

Purpose

To achieve automated identification of lesions in Optical Coherence Tomography images based on a lightweight convolutional neural network architecture.

Methods

This retrospective study developed a lightweight lesion recognition and classification model based on deep learning, utilizing OCT images from Sichuan Provincial People’s Hospital as well as datasets including OCT-C8, OCT-2017, and HD-OCT of MH. A total of 574,808 training images and 71,851 validation images were included. Model performance evaluation was conducted in two phases: First, the model’s generalization capability and stability were tested on four independent external validation sets (with image counts of 71,854, 2,763, 1,650, and 785, respectively). Second, the clinical experience level corresponding to the model’s diagnostic proficiency was assessed by comparing its classification performance with that of ophthalmologists at different stages of clinical expertise.

Results

The model achieved an average accuracy of 87.40%, an F1-score of 75.78%, and a ROC_AUC value of 98.68% on the validation set. For the nine categories of OCT images—neovascularization, posterior vitreous detachment, epiretinal membrane, macular hole, macular schisis, subretinal fluid, normal images, vitreomacular traction, and obscured images—the model’s specificity was 99.42%, 98.20%, 95.43%, 99.01%, 96.97%, 99.66%, 99.15%, 99.54%, and 98.56%, respectively; while its recall rates were 94.29%, 74.16%, 97.94%, 76.17%, 89.16%, 84.03%, 73.08%, 73.23%, and 98.44%, respectively. The model achieved an average accuracy of 89.42% in the external validation sets. In comparison, junior physicians had an accuracy of 74% (95% CI: 61.84%–86.16%) in interpreting OCT images, while mid-career physicians had an accuracy of 88% (95% CI: 78.99%–97.01%). These results indicate that the model’s performance approaches that of mid-career physicians.

Conclusion

This study achieved automated identification of nine categories of OCT images, with its clinical performance initially reaching the level of mid-career physicians. Moreover, the model can be deployed locally in medical environments to enhance diagnostic efficiency and accuracy.

Introduction

Globally, retinal diseases are a major cause of vision loss and blindness, with common conditions including macular degeneration and diabetic retinopathy, among others [1]. These diseases significantly impact the quality of life for millions of people. Optical Coherence Tomography (OCT), as a non-invasive and high-resolution medical imaging technique [2], is widely applied in ophthalmology, playing a crucial role in the diagnosis, treatment, and follow-up of retinal diseases. OCT provides micron-level images of retinal structures, enabling clinicians to clearly observe subtle lesions [3], such as choroidal neovascularization, macular schisis, and subretinal fluid. Beyond detecting subtle lesions, OCT also assists doctors in assessing retinal thickness and other pathological changes [4], monitoring the progression of retinal diseases and the effectiveness of treatments, thereby facilitating personalized treatment plans.

Although OCT can provide detailed anatomical structural information, the analysis of its images still heavily relies on the personal experience of ophthalmologists, leading to a cumbersome diagnostic process that is prone to subjectivity [5]. Additionally, different doctors may arrive at varying diagnostic conclusions for the same image due to differences in experience, expertise, and contextual factors [6], which can affect diagnostic consistency and accuracy while increasing the risk of misdiagnosis [7]. Therefore, research into objective and accurate automated diagnostic methods is of great importance. In recent years, deep learning technologies have made significant progress in the field of medical image analysis, particularly with Convolutional Neural Networks (CNNs) demonstrating excellent performance in image classification and object detection [8]. However, traditional deep learning architectures are often computationally complex and involve a large number of parameters [9], limiting their efficient application in clinical settings. As a result, lightweight neural networks have emerged, which reduce model parameters and computational complexity while maintaining high accuracy. This enables faster detection speed, assisting doctors in making rapid diagnostic decisions [10], and making them particularly suitable for medical environments with limited computational resources. Furthermore, their lightweight and computationally efficient characteristics effectively lower the barriers to hardware deployment and reduce operational maintenance costs [11].

Pang et al. [12] achieved 98.1% accuracy, 97.9% sensitivity, and 98.0% specificity in classification tasks for CNV, DURSEN, DME, and NORMAL images using the MSA-NET model. Wang et al. [13] utilized the EfficientNet model to train on CNV, DME, DRUSEN, and NORMAL images, attaining an accuracy of 99.2%. Elkholy et al. [14] employed the VGG16 model to classify 35,468 OCT images, covering normal retina, diabetic macular edema, choroidal neovascularization, and age-related macular degeneration, ultimately achieving an accuracy of 97%.

Although existing methods demonstrate high accuracy in classifying specific lesions, their models are often computationally intensive, have limited category coverage, and require high deployment thresholds. These limitations hinder their widespread adoption in primary healthcare settings. Unlike the aforementioned studies, this research employs a lightweight neural network combined with a semi-supervised learning strategy to achieve intelligent recognition of nine types of OCT image lesions. This approach addresses the challenges of deploying models in resource-constrained primary healthcare environments while reducing the risks of misdiagnosis and missed diagnosis caused by subjective physician judgment. Consequently, it provides patients with more precise auxiliary diagnostic support and enhances overall clinical treatment outcomes.

Methods

Hardware configuration

The hardware configuration used in the experiments includes an i5 12600KF CPU and an NVIDIA RTX 4060ti GPU, with 32GB of RAM. The deep learning framework employed is PyTorch 2.3.1, utilizing CUDA 11.8, and Python 3.9 is used as the programming language. Data analysis is supported by scientific computing libraries such as Scipy and Statsmodels. All experiments are conducted in a CUDA runtime environment.

Datasets

This study has obtained ethical approval from the Ethics Review Committee of Sichuan Provincial People’s Hospital [Ethics Review (Research) No. 794, 2025] and complies with the relevant provisions of the Declaration of Helsinki. As a retrospective diagnostic trial, patient data were anonymized for analysis, exempting the requirement for informed consent. The study utilized OCT images from Sichuan Provincial People’s Hospital collected between June 2023 and October 2024, as well as the OCT-C8 dataset, OCT-2017 dataset, and HD-OCT of MH dataset. Six resident physicians were invited to classify lesions based on OCT images, covering nine categories: normal OCT images, vitreomacular traction, choroidal neovascularization, subretinal fluid, macular schisis, epiretinal membrane, retinal hole, posterior vitreous detachment, and obscured images. To address common occlusion artifacts in OCT imaging, this study designed a data augmentation method based on physical characteristics. Specifically, during the image preprocessing stage, continuous regions along the X-axis were randomly selected from normal OCT images. The RGB values of all Y-axis pixels within the selected region were set to (0,0,0) to simulate complete occlusion. Subsequently, random white-point noise (noise intensity: 0.3) was superimposed to mimic the typical occlusion features caused by poor probe contact or abnormal tissue reflections in real OCT devices. The primary purpose of constructing the obscured image dataset was to exclude interfering images during OCT lesion recognition, thereby improving the recognition accuracy for other lesions.

This study employed a standardized image preprocessing pipeline, resizing all input images to a uniform resolution of 224 × 224 pixels to ensure compatibility with the input requirements of mainstream deep convolutional neural networks. During the data augmentation phase, a systematic multi-scale data augmentation strategy was implemented based on the imaging characteristics of optical coherence tomography (OCT) images. Specific operations included a 50% probability of random horizontal flipping, random translation within a range of [-30, 30] pixels along the XY axes, and random rotation within a range of [-30, 30] degrees. Subsequently, anisotropic scaling was applied while preserving the original aspect ratio, with random compression ratios between 0.7 and 0.9 along the X-axis. This series of data augmentation operations effectively simulated various imaging variations encountered in clinical practice, such as differences in scanning angles, variations in patient positioning, and fluctuations in device parameters. Through this strategy, the original dataset was expanded to include 718,513 OCT images with rich morphological variations. For dataset partitioning, stratified random sampling was used to divide the data into training (574,808 images), validation (71,851 images), and test (71,854 images) sets in an 8:1:1 ratio.

To rigorously evaluate the model’s robustness to morphological changes in lesions and prevent overfitting between the validation set and training set, this study implemented a systematic data augmentation strategy on the validation set: horizontal flipping and XY-axis translation were performed sequentially to simulate image variations caused by equipment differences and positional changes; subsequently, anisotropic scaling was applied while preserving the original aspect ratio, with compression along the X-axis to 0.8 times the original size and expansion along the Y-axis to 1.2 times the original size, simulating pathological deformations during lesion progression; finally, Gaussian noise with a density of 0.06 was added to replicate signal interference during image acquisition.

Overview of architecture

This study constructed a nine-class OCT image lesion classification model based on the lightweight MobileNet convolutional neural network architecture (Fig. 1A), providing feasibility for localized deployment on clinical terminal devices. MobileNet employs depthwise separable convolution as its core design, decomposing standard convolution into depthwise convolution and pointwise convolution [15]. Combined with inverted residual blocks and linear bottleneck structures [16], this approach significantly reduces computational complexity while maintaining model accuracy. The architecture enhances feature expression through nonlinear activation functions and leverages residual connections to mitigate gradient vanishing issues [17], making it particularly suitable for application scenarios such as medical image processing, where a balance between accuracy and efficiency is required [18]. In the experiments, a transfer learning strategy was adopted. The model was initialized using ImageNet pre-trained weights, with input images resized to 224 × 224 pixels. The Adam optimizer (learning rate: 1e-4, weight decay: 1e-5) was used for training, with a batch size of 16 and a random seed (seed = 42). The model was trained for 10 epochs, and its performance was monitored using an independent validation set (primarily including metrics such as accuracy, recall, F1-score, and AUC). The model weights achieving the highest accuracy on the validation set throughout the entire training process were selected as the final model. Additionally, the performance differences of the ResNet architecture on this dataset were compared.

Fig. 1.

Network Structure Diagram of the OCT Image Lesion Recognition Model. (A) Schematic diagram of the MobileNet network structure. (B) Construction of the Eye_OCT database and the model training workflow: Six resident physicians classified images according to standardized criteria, establishing the Eye_OCT dataset comprising nine categories. Model training was conducted on this dataset based on the MobileNet architecture, with performance comparison and analysis against the ResNet architecture. Furthermore, this framework synchronizes the prediction results of each image recognition in real-time to the semi-supervised learning database. Abbreviations: RT: Retinal Tear PVD: Posterior Vitreous Detachment VMT: Vitreomacular Traction CNV: Choroidal Neovascularization SRF: Subretinal Fluid ERM: Epiretinal Membrane Normal: Normal Retina RS: Retinal Schisis Blurry: Blurry Image

To enhance the model’s ability to recognize lesion features, this study innovatively proposes a semi-supervised learning framework (Fig. 1B). Its core lies in establishing a high-confidence pseudo-label generation strategy with a hierarchical quality control mechanism for screening unlabeled data. To address the issue of class imbalance in the dataset, a dynamic threshold adjustment method is employed. For long-tail categories with scarce samples, the confidence threshold is actively lowered to expand the number of qualified pseudo-labels, while for conventional categories with sufficient samples, a strict confidence standard of 0.95 or higher is maintained. To ensure the quality of pseudo-labels, a hierarchical quality assurance mechanism is established. All samples from long-tail categories that pass the dynamic threshold screening undergo full review by clinical doctors, while samples from conventional categories are subjected to random sampling for review. This strategy effectively controls pseudo-label errors through high-confidence screening, enhancing the credibility of positive detection results and thereby improving the model’s overall prediction accuracy and generalization ability. Ultimately, all pseudo-label data that pass the review will form a “candidate training set.” After a secondary review by the research team, it will be merged into the training set for the deep training and optimization of the lesion diagnosis model, with a focus on improving the recognition of rare lesions. Simultaneously, this database will serve as a foundational data resource for subsequent algorithm iterations and advanced research, continuously supporting the formation of a closed-loop research process.

Statistical analysis

To evaluate the performance of the OCT lesion classification model in this study, systematic statistical analysis was first conducted on the model’s performance during the training process. Multiple evaluation metrics, including accuracy, recall, specificity, precision, F1-score, and the ROC_AUC curve, were employed to quantify the model’s performance. The specific calculation formulas are shown below:

Secondly, to evaluate the generalizability and accuracy of the proposed OCT image lesion recognition model, this study validated it on four independent external datasets, including: an OCT image dataset provided by the Sichuan Provincial Geriatrics Center (Fig. 2A), the publicly available OCT_Dataset_Rabbani (Fig. 2B) and OCT_Dataset_Srinivasan (Fig. 2C) from the Kaggle platform, and the Eye_OCTEval test set constructed in this study (Fig. 2D). Based on the original OCT images, the test set underwent sequential horizontal flipping, XY-axis translation, anisotropic scaling (X-axis compressed to 0.8 times, Y-axis expanded to 1.2 times), and Gaussian noise enhancement with a noise density of 0.07, aiming to generate more challenging morphological variations and noisy images to comprehensively test the model’s performance under complex conditions. The four external datasets involved a total of three resident physicians classifying the images based on lesion features and excluding blurred OCT images. The classification results were submitted to a chief physician for review and confirmation to avoid classification errors by the resident physicians. Simultaneously, this study categorized physicians from the Retina Research Group of Sichuan Provincial People’s Hospital into three groups based on clinical experience levels: junior group (1–2 years of clinical experience), mid-career group (3–5 years of clinical experience), and senior group (attending physician or higher), with two participants in each group jointly involved in lesion identification. The accuracy of their judgments was evaluated by calculating the 95% confidence interval (two-sided test, α = 0.05), and all evaluated images had all potentially influencing label information concealed. During the analysis phase, an independent analyst performed blinded processing on all evaluation results, ensuring no knowledge of the specific source of the results. Finally, the analyst compared the model’s classification and diagnostic outcomes with those of the junior, mid-career, and senior ophthalmologist groups to determine which clinical experience level the model’s diagnostic capability corresponds to.

Fig. 2.

Sample Size Distribution of the Four External Validation Datasets for the OCT Image Lesion Recognition Model. (A) Distribution of the Eye_OCTEval dataset, comprising a total of 71,854 images. (B) Distribution of the Sichuan Provincial Geriatrics Center dataset, comprising a total of 2,763 images. (C) Distribution of the OCT_Dataset_Rabbani dataset, comprising a total of 1,650 images. (D) Distribution of the OCT_Dataset_Srinivasan dataset, comprising a total of 785 images

Results

This study conducted a systematic performance evaluation and comparative analysis of two network architectures, MobileNet and ResNet, for the classification task of nine types of OCT images on the validation set (Figs. 3A and B). Although the OCT lesion diagnosis model developed based on the lightweight MobileNet architecture exhibited slightly lower accuracy (Difference: -0.71%) and precision (Difference: -10.74%) compared to the ResNet model (Table 1), its model size was only 6.6% of that of ResNet, with a 0.75 ms improvement in processing a single image, demonstrating significantly lower parameter count and higher computational efficiency (Table 1). The model achieved accuracy rates of 87.57%, 84.79%, 92.34%, and 92.97% on the Eye_OCTEval, Sichuan Provincial Geriatrics Center, OCT_Dataset_Rabbani, and OCT_Dataset_Srinivasan datasets, respectively; precision rates of 81.24%, 74.25%, 91.87%, and 93.38%; and specificity rates of 98.27%, 95.30%, 97.35%, and 96.37%. For the nine categories of OCT images—neovascularization, posterior vitreous detachment, epiretinal membrane, macular hole, macular schisis, subretinal fluid, normal images, vitreomacular traction, and obscured images—the model’s specificity rates were 99.42%, 98.20%, 95.43%, 99.01%, 96.97%, 99.66%, 99.15%, 99.54%, and 98.56%, respectively; while its recall rates were 94.29%, 74.16%, 97.94%, 76.17%, 89.16%, 84.03%, 73.08%, 73.23%, and 98.44%, respectively.

Fig. 3.

Performance evaluation and visualization results of the lesion recognition model. (A) Comparative analysis of comprehensive metrics for OCT image lesion recognition between the two models. (B) Detailed performance metrics of the proposed lesion recognition model in this study. (C) Heatmap visualization of key features for nine types of lesions identified by the OCT image lesion recognition model. The blue regions in the images represent the lesion feature areas focused on by the model. All images are selected from the Eye_OCT dataset. To ensure strict patient privacy protection, all personal information in this dataset was anonymized during the collection process: (a) normal image; (b) posterior vitreous detachment; (c) epiretinal membrane; (d) retinal tear; (e) vitreomacular traction syndrome; (f) neovascularization; (g) retinal schisis; (h) subretinal fluid; (i) occluded image. (D) Confusion matrices of the four external validation sets. (a) Confusion matrix of the model on the external validation set from the Sichuan Provincial Geriatrics Center, showing that the lesion recognition model is insufficient in distinguishing between epiretinal membrane and normal images. (b) Confusion matrix of the model on the external validation set from OCTEval, indicating that the model lacks effectiveness in distinguishing between posterior vitreous detachment and normal images. (c) Confusion matrix of the model on the external validation set from OCT_Dataset_Rabbani. (d) Confusion matrix of the model on the external validation set from OCT_Dataset_Srinivasan

Table 1.

Performance comparison between lightweight neural networks and traditional large neural networks. While MobileNet slightly underperformed ResNet in the nine-class OCT lesion recognition task, its model size is only 6.6% of ResNet’s, and its processing time per image is improved by 0.75 Ms

Performance comparison of the two models in OCT image lesion recognition
Model / Metric	Accuracy	F1	Recall	ROC_AUC	Precision	Specificity	Size	Time
MobileNet	87.40%	75.78%	84.50%	98.68%	73.08%	98.47%	5960KB	2.49 ms
ResNet	88.11%	80.13%	78.28%	97.84%	83.82%	98.33%	90100KB	3.24 ms

The OCT lesion recognition model in this study was evaluated on four external validation datasets that were not involved in the model training. The average accuracy across the four external validation datasets was 89.42%, with an average specificity of 96.82%, an average precision of 85.19%, and an average Kappa value of 81.98%. The model achieved an average accuracy of 89.42% on the external validation sets. In comparison, junior physicians had an accuracy of 74% (95% CI: 61.84%–86.16%) in interpreting OCT images, while mid-career physicians had an accuracy of 88% (95% CI: 78.99%–97.01%). These results indicate that the model’s performance approaches that of mid-career physicians.

The OCT lesion recognition model proposed in this study (Fig. 3C) demonstrated high recognition precision across multiple validations. However, confusion matrix analysis (Fig. 3D) revealed that the model still has certain limitations in the following differential diagnoses: (1) distinguishing between normal fundus and posterior vitreous detachment, and (2) differentiating normal fundus from epiretinal membrane.

Discussion

Compared to studies by Pang [12], Wang [13], and Elkholy [14], which involved tens of thousands of images and focused on only four types of lesions, this research builds a lightweight lesion recognition model using over 700,000 OCT images (approximately 140,000 original images). It covers the nine most common types of OCT images encountered in primary healthcare institutions, achieving significant breakthroughs in both disease coverage and data scale. Double-blind tests confirmed that its recognition accuracy reaches the level of mid-career physicians (3–5 years of clinical experience). Furthermore, the model’s size is only 6.6% of that of traditional large neural networks, with a processing speed increase of 0.75 ms per image. The lightweight network structure effectively lowers hardware deployment barriers and reduces operational maintenance costs. Finally, the model possesses semi-supervised learning capabilities, enabling continuous learning and performance optimization from new data.

The lightweight model for OCT lesion diagnosis developed in this study, compared to traditional large neural networks [19], not only improves the response speed of real-time inference but also reduces computational resource consumption. Adopting a lightweight design, this model can efficiently and stably perform OCT lesion diagnosis in environments with limited hardware capabilities and achieve low-cost localized deployment. This makes it particularly suitable for primary hospitals with relatively constrained medical resources, as it avoids the risk of data leakage associated with cloud data transmission [20] and enhances accessibility in remote areas with poor network conditions. More importantly, compared to traditional methods relying on physicians’ expertise for lesion identification in images [21], this model consistently adheres to the same diagnostic criteria for lesions, effectively reducing interference from subjective factors among physicians. This improves diagnostic consistency and reliability while lowering the risk of misdiagnosis due to differences in physicians’ experience. Additionally, the lesion diagnostic performance of this model is already close to that of mid-career physicians, enabling it to perform automated lesion screening in primary community hospitals. This promotes early detection and intervention of ophthalmic diseases such as diabetic retinopathy, holding significant clinical value for implementing hierarchical diagnosis and treatment and enhancing public health service standards.

Although the OCT lesion diagnosis model developed in this study demonstrates promising clinical potential, it still has certain limitations. On the technical level, the model’s ability to identify “weak-feature” lesions requires improvement: (1) The subtle separation of the posterior vitreous membrane in early-stage posterior vitreous detachment (PVD) is easily confused with its normal adherent state in OCT images; (2) Early-stage epiretinal membrane (ERM) manifests only as enhanced reflectivity of the inner retinal surface and subtle textural changes, with no significant morphological alterations. Both types of early-stage lesions, due to their concealed features and subtle morphological evolution, make it difficult for the model to extract effective diagnostic cues from OCT images with low contrast and blurred edges, ultimately leading to higher misclassification rates. On the clinical application level, the diagnostic performance of the model may be influenced by multiple factors: (1) The model’s performance may vary across different patient populations and pathological characteristics, particularly in complex cases where certain lesion features may be challenging to accurately identify, leading to misdiagnosis or missed diagnosis [22]. However, this is a common challenge in the field of image recognition research. (2) The performance of deep learning models heavily depends on the quality and diversity of the training dataset. Although this study successfully constructed the EyeOCT database, which encompasses a vast amount of OCT lesion data, the high diversity and clinical complexity of ophthalmic disease manifestations prevent the database from covering all possible clinical scenarios and lesion patterns. This may result in insufficient diagnostic accuracy for certain specific cases, and there may be some uncertainty in diagnostic outcomes, particularly when dealing with rare or complex lesions not included in the database.

Future research should focus on deeper integration of this model with ophthalmic clinical practice, particularly in two key areas: data diversity and multimodal data fusion. Firstly, collecting larger and more diverse clinical datasets—including data from different populations, pathological characteristics, and device types [23]—will enhance the model’s generalizability and improve its adaptability across various clinical settings. At the same time, the application of lightweight models in ophthalmology is particularly important, as they can be efficiently deployed in resource-constrained environments, reduce diagnostic costs, and increase accessibility in remote areas. Additionally, research should further explore multimodal data fusion [24], especially the integration of OCT with other medical data. This will improve the accuracy of early disease screening and provide more precise support for personalized treatment decisions.

Conclusion

This study achieved automated recognition of nine categories of OCT images, with its clinical performance approaching the level of mid-career physicians. Moreover, the model can be deployed locally in medical environments to enhance diagnostic efficiency and accuracy.

Author contributions

A.B. C.D. E.Fwrote the main manuscript text and G.H. I.Hprepared Figs. 1, 2 and 3. All authors reviewed the manuscript.

Funding

National Natural Science Foundation of China (Grant No. 62376173).

Data availability

The datasets used and supporting the conclusions of this study are available from the corresponding author upon reasonable request, in compliance with ethical approval, data protection agreements, and relevant laws and regulations.

Declarations

Ethical approval and consent to participate

This study has obtained ethical approval from the Review Committee of Sichuan Provincial People’s Hospital [Approval No. 794, 2025] and complies with the relevant provisions of the Declaration of Helsinki. As a retrospective diagnostic trial, patient data were analyzed anonymously, and informed consent was waived.

Consent for publication

Not applicable. This study did not use any personally identifiable data, images, or detailed information.

Competing interests

The authors declare no competing interests.