KD-OCT: Knowledge Distillation Enables Efficient Retinal OCT Classification with Lightweight EfficientNet-B2 Models

Age-related macular degeneration and related conditions represent a major global cause of vision loss, with optical coherence tomography (OCT) playing a crucial role in early diagnosis and treatment. However, deploying advanced deep learning models for real-time clinical use presents significant computational challenges. To address this, Erfan Nourbakhsh from University of Isfahan, Nasrin Sanjari from Shahid Beheshti University of Medical Science, and Ali Nourbakhsh from Isfahan University of Technology, alongside their colleagues, developed KD-OCT, a novel knowledge distillation framework. This method successfully compresses a powerful, high-performance model into a much smaller, more efficient network without sacrificing diagnostic accuracy, paving the way for widespread, accessible, and rapid AMD screening even on resource-constrained devices.

The team demonstrates that KD-OCT achieves near-expert performance while substantially reducing model size and inference time, exceeding the capabilities of existing OCT classification frameworks. However, deploying advanced deep learning models for real-time clinical use presents significant computational challenges. To address this, researchers developed KD-OCT, a novel knowledge distillation framework that compresses a powerful, high-performance model into a much smaller, more efficient network without sacrificing diagnostic accuracy. This advancement paves the way for widespread, accessible, and rapid AMD screening even on resource-constrained devices.,.

Knowledge Distillation Improves AMD Detection Efficiency This research presents a novel approach to detecting Age-Related Macular Degeneration (AMD) from Optical Coherence Tomography (OCT) images using knowledge distillation. The goal is to create a model that is both accurate and efficient, addressing the need for reliable AMD diagnosis in clinical settings. Researchers leverage knowledge distillation to transfer knowledge from a large, accurate model to a smaller, more efficient one, allowing the student model to achieve comparable performance while requiring significantly fewer computational resources.

The team utilized a labeled dataset of retinal OCT images, categorized into normal, drusen, and choroidal neovascularization cases. The proposed knowledge distillation approach demonstrates promising results, achieving high accuracy in AMD detection while significantly reducing model size and computational cost, making it suitable for deployment in real-world clinical settings.,. Real-time Knowledge Distillation for Retinal Image Classification This study pioneers KD-OCT, a novel knowledge distillation framework designed to compress a high-performance ConvNeXtV2-Large model into a lightweight EfficientNet-B2 model for classifying retinal OCT images as normal, exhibiting drusen, or indicating choroidal neovascularization.

The team enhanced the teacher model with advanced data augmentations and stochastic weight averaging to maximize its performance before distillation. KD-OCT employs a real-time distillation process, balancing soft knowledge transfer from the teacher with hard ground-truth supervision, to ensure the student model learns effectively. Scientists utilized the Noor Eye Hospital dataset and implemented a patient-level 5-fold cross-validation scheme to rigorously evaluate the framework. Experimental results demonstrate that KD-OCT significantly outperforms comparable multi-scale and feature-fusion OCT classifiers in terms of efficiency and accuracy. The compressed student model achieves near-teacher performance despite containing 25. 5times fewer parameters, representing a substantial reduction in computational demands and facilitating potential edge deployment for real-time AMD screening.,. Efficient AMD Screening via Knowledge Distillation This work presents KD-OCT, a novel knowledge distillation framework designed to compress a high-performance ConvNeXtV2-Large model into a lightweight EfficientNet-B2 model for classifying retinal OCT images.

The team successfully achieved substantial model compression while maintaining diagnostic accuracy, addressing a critical need for efficient, real-time AMD screening tools. KD-OCT employs a real-time distillation process, balancing soft knowledge transfer from the teacher model with direct supervision from ground-truth labels, and was rigorously evaluated using patient-level 5-fold cross-validation on the Noor Eye Hospital dataset. Experimental results demonstrate that the compressed student model achieves near-teacher performance, surpassing comparable multi-scale and feature-fusion OCT classifiers in terms of efficiency and accuracy. Specifically, the EfficientNet-B2 student model contains 25. 5times fewer parameters than the original ConvNeXtV2-Large teacher, significantly reducing computational demands without compromising diagnostic capability, enabling potential deployment on edge devices and facilitating wider access to automated AMD screening.,. Knowledge Distillation for Efficient Retinal Diagnosis This study presents KD-OCT, a new knowledge distillation framework designed to compress complex deep learning models for diagnosing age-related macular degeneration and related conditions from retinal optical coherence tomography (OCT) images. Researchers successfully transferred the knowledge from a large, high-performing model, ConvNeXtV2-Large, into a smaller, more efficient EfficientNet-B2 model, achieving comparable accuracy with a substantial reduction in computational demands. The method employs a real-time distillation process, balancing the transfer of soft knowledge from the teacher model with direct supervision from ground-truth data, and demonstrates superior performance compared to existing multi-scale and feature-fusion approaches. Experimental results, validated using patient-level cross-validation on multiple datasets, show that KD-OCT maintains high diagnostic accuracy, reaching 92-98%, while reducing the number of parameters by a factor of 25. 5 and accelerating inference speed. This compression facilitates the deployment of accurate diagnostic tools on resource-constrained platforms, such as portable devices, and promotes scalable AMD screening programs. Researchers acknowledge that stochastic weight averaging caused a slight performance decrease, but highlight the significant benefit of focal loss in addressing class imbalance, particularly in identifying subtle cases of choroidal neovascularization. Future research will focus on reducing the reliance on labeled data through semi-supervised learning, exploring the integration of fundus images to further improve accuracy, and extending the framework to diagnose other retinal pathologies like diabetic macular edema. These advancements aim to optimize the system for seamless integration into portable diagnostic devices, ultimately enhancing accessibility and improving patient care. 👉 More information 🗞 KD-OCT: Efficient Knowledge Distillation for Clinical-Grade Retinal OCT Classification 🧠 ArXiv: https://arxiv.org/abs/2512.09069 Tags: