Highlights<\/h2>\n\n\n\n\n- Novel energy based early exiting schema.<\/li>\n\n\n\n
- Novel method of early exit detection using distribution fitting.<\/li>\n\n\n\n
- Real-time OCT classification on low computation devices.<\/li>\n<\/ul>\n\n\n\n
Abstract<\/h2>\n\n\n\n
Background:<\/p>\n\n\n\n
The increasing prevalence of retinal diseases highlights the need for more ophthalmology specialists, as existing diagnostic methods struggle to efficiently screen a growing patient population. Recent advancements in artificial intelligence, particularly in deep learning, offer a promising solution.<\/p>\n\n\n\n
Aims:<\/h3>\n\n\n\n
In light of the rising demand for telehealth services and precise diagnostic systems, we aim to develop a deep-learning framework for classifying Optical Coherence Tomography (OCT) images. This framework is designed to be used by ophthalmologists on cost-effective, low-computation devices such as smartphones.<\/p>\n\n\n\n
Materials and Methods:<\/h3>\n\n\n\n
We present an energy-based early-exiting deep learning framework to optimize computational efficiency for OCT classification. This framework integrates an early exit module along with a distribution-based decision-making algorithm for inference time. The modules will raise exit numbers from earlier layers of the backbone network.<\/p>\n\n\n\n
Results:<\/h3>\n\n\n\n
Our ResNet-18 and ResNet-50 models achieved an impressive 100% accuracy in under 0.1 s on the OCT2017 dataset, averaging 0.77 GMacs per sample. In testing on a second dataset comprising 3,231 samples, we maintained an F-score of 1 while successfully reducing computational overhead. Furthermore, the model exhibited excellent performance on smartphones, averaging 0.618 s per sample with just 6 MB of memory usage.<\/p>\n\n\n\n
Conclusion:<\/h3>\n\n\n\n
We have developed deep-learning models that enhance inference speed, reduce computational costs, and maintain high accuracy. This framework has the potential to support ophthalmologists in diagnosing retinal diseases using smartphones or serve as a valuable telehealth monitoring tool.<\/p>\n\n\n\n
Graphical abstract<\/h2>\n\n\n\n![\"\"\/](\"https:\/\/ars.els-cdn.com\/content\/image\/1-s2.0-S174680942401423X-ga1.jpg\")
<\/a><\/figure>\n\n\n\n
Abstract<\/h2>\n\n\n\n
Background:<\/p>\n\n\n\n
The increasing prevalence of retinal diseases highlights the need for more ophthalmology specialists, as existing diagnostic methods struggle to efficiently screen a growing patient population. Recent advancements in artificial intelligence, particularly in deep learning, offer a promising solution.<\/p>\n\n\n\n
Aims:<\/h3>\n\n\n\n
In light of the rising demand for telehealth services and precise diagnostic systems, we aim to develop a deep-learning framework for classifying Optical Coherence Tomography (OCT) images. This framework is designed to be used by ophthalmologists on cost-effective, low-computation devices such as smartphones.<\/p>\n\n\n\n
Materials and Methods:<\/h3>\n\n\n\n
We present an energy-based early-exiting deep learning framework to optimize computational efficiency for OCT classification. This framework integrates an early exit module along with a distribution-based decision-making algorithm for inference time. The modules will raise exit numbers from earlier layers of the backbone network.<\/p>\n\n\n\n
Results:<\/h3>\n\n\n\n
Our ResNet-18 and ResNet-50 models achieved an impressive 100% accuracy in under 0.1 s on the OCT2017 dataset, averaging 0.77 GMacs per sample. In testing on a second dataset comprising 3,231 samples, we maintained an F-score of 1 while successfully reducing computational overhead. Furthermore, the model exhibited excellent performance on smartphones, averaging 0.618 s per sample with just 6 MB of memory usage.<\/p>\n\n\n\n
Conclusion:<\/h3>\n\n\n\n
We have developed deep-learning models that enhance inference speed, reduce computational costs, and maintain high accuracy. This framework has the potential to support ophthalmologists in diagnosing retinal diseases using smartphones or serve as a valuable telehealth monitoring tool.<\/p>\n\n\n\n
Graphical abstract<\/h2>\n\n\n\n<\/a><\/figure>\n\n\n\n
<\/a><\/figure>\n\n\n\n