Ophthalmology evolution has long been awaited at public hospitals common in the United States, United Kingdom, Australia, Singapore and Hong Kong. A study in the United Kingdom found that 22 weeks of delay in treatment resulted in patients suffering harm that could have been avoided with previous interventions, such as permanent deterioration of vision and loss of vision. This reflects the urgent public health need for new solutions to make primary, secondary and tertiary ophthalmic medical services available to more people.
How does AI affect ophthalmic diagnosis and screening?
Despite these research reports on excellent DL performance in clinical validation studies, few studies have evaluated a format suitable for practical implementation. It can be either a “fully automated” model or a “semi-automated” model. A “fully automated model” is a human provider by having an AI system initiate referrals to ophthalmologists as needed and flag patients for continuous community-based monitoring. Works without the involvement of. In contrast, “semi-automatic models” can take various forms involving human evaluators or ophthalmologists to extend DL classification as a tool for patient triage.
Xie et al. I explained the cost-effectiveness of such a semi-automatic model developed in Singapore. This model uses AI to triage the CFP, and the model flagged as anomalous is reviewed by a human evaluator using remote ophthalmology. In addition to determining the ideal form of implementation, there are many other practical, technical and sociocultural challenges to overcome.
What are the current challenges in implementing AI in clinical practice?
Practical Challenges The first major challenge in implementing a validated AI solution for
Ophthalmology evolution is the need for a working overall solution. This may require a combination of DL systems with clinically acceptable performance and solution interoperability to receive different quality images from commonly used devices. Supportive clinical guidelines for clinical communication and patient selection for patients in the DL system classification should also be developed. Grasman et al. Previously, the importance of proper patient selection in reports of misclassification by the DL system caused by the macular reflex in young patients in the validation dataset when the boy was not represented in the DL system training dataset was shown.
Technical Challenges
First, the need for proper training data and external validation is a major technical challenge that needs to be addressed to facilitate the generalization and translation of these Ophthalmology evolution solutions. This was reported by Goh et al. I explained it in more detail in her article that caused this problem. This also poses the following challenges: It is a tedious labeling of the input data of the training process and requires the involvement of experienced practitioners who make human error. Repeated need to label image datasets to calibrate the DL system for each new population, coupled with the need for careful patient selection, can delay recruitment and increase setup costs.
However, recent research has used data from detailed 3D scans such as OCT to train DL algorithms to automate the labeling of training data and to do other types of scans such as Classify2. It has been shown that you can train a new DL algorithm to run. -Dimensional CFP. This helps scale up the calibration process to address human limitations and improve performance when applying the DL system to a new target population.
Read my more blogs from here
