计算机科学
可解释性
人工智能
光学相干层析成像
卷积神经网络
计算机视觉
稳健性(进化)
变压器
模式识别(心理学)
可扩展性
医学影像学
可视化
深度学习
训练集
图像质量
平滑的
量化(信号处理)
机器学习
边缘检测
边缘设备
人工神经网络
视网膜
监督学习
机器视觉
作者
Jesvanth S,Karan Kumar K,Devadharshini S.,S Sudharson
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2025-12-19
卷期号:: 1-8
标识
DOI:10.1109/cict67193.2025.11398870
摘要
The transformer-based automated model to classify retinal diseases using Optical Coherence Tomography (OCT) images and specifically aiming to be deployed in edge devices. The Convolutional Neural Networks (CNNs) and other traditional methods require significant amount of annotated data and computing power and therefore can only be used in real-time or resource-poor settings such as mobile diagnostic devices. To overcome these limitations, this paper suggests a lightweight Vision Transformer (ViT) that will be pretrained through the self-supervised learning approach namely, Masked Autoencoders (MAE) to useful utilize unlabelled OCT data. To build robustness to sub-optimal imaging conditions the scores assigned to image quality are also integrated directly into the training pipeline. Additionally, quantization and pruning strategies has been adopted to compress the model to fit well in edge-deployment. Proposed model is validated by performing scalable experiments on publicly accessible datasets, such as OCT2017 and OCTDL, and outperforms the traditional CNN-based baselines in accuracy and speed of the inference, and model size. Moreover, the attention rollout visualizations are used to increase clinical interpretability of the system making it practically efficient and explainable in the real-world ophthalmic setting.
科研通智能强力驱动
Strongly Powered by AbleSci AI