A Lightweight OCT Image Classification Model with Low Configuration and High Efficiency

The existing Retina OCT image automatic classification systems encounter challenges in deployment due to their substantial size. To address this, we propose Light-AP-EfficientNet, a lightweight model leveraging adaptive pooling for efficient feature extraction and classification, specifocally designed for effective data mining applications in medical imaging. Firstly, we optimize EfficientNet’s convolutional layer settings to reduce redundant convolutional structures, significantly reducing the model’s parameters. Then, we integrate adaptive pooling layers to facilitate the model in learning both global and local features, enhancing model classification performance. Experimental results demonstrate that Light-AP-EfficientNet achieves 99.7% accuracy on UCSD dataset, while requiring only 17% of the parameter volume of ShuffleNetV2 and 19% of the computational volume of MobileNetV2. Additionally, it processes a single image in just 0.028 s on a CPU and 0.009 s on a GPU. Furthermore, compared to recent novel models, our model demonstrates significant improvements in metrics such as Accuracy and Precision on the same dataset. Specifically, the maximum improvement in Accuracy is 4.5%, and in Precision is 5.42%. With its high accuracy and reduced hardware requirements, Light-AP-EfficientNet is ideal for data mining tasks in resource-constrained scenarios.