Revolutionizing diabetic retinopathy screening and management: The role of artificial intelligence and machine learning

Table 1 Overview of diabetic retinopathy diagnostic tools

Tool	Year introduced	Country of origin	Ref.	Advantages	Disadvantages	AI/ML-based
Fundus photography	Mid-20th century	Germany	Srinivasan et al[13], 2023	Established method for capturing detailed retinal images	Resource-intensive requires specialized personnel, expensive, and not scalable in low-resource settings	No
Optical coherence tomography	1991	United States	Huang et al[24], 1991	High-resolution cross-sectional images; effective in detecting diabetic macular edema.	High cost, requires specialized training, limited availability in low-resource settings	No
Fluorescein angiography	1961	United States	Norton and Gutman[27], 1965	Gold standard for visualizing retinal vasculature; highly precise.	Invasive, requires dye injection, potential side effects, limited use in rural and low-income areas.	No
Ultrawide-field imaging	Early 2000s	Canada, United Kingdom	Nagiel et al[32], 2016	Captures up to 200 degrees of the retina; detects peripheral lesions often missed by standard imaging	High cost, requires specialized training, limited adoption in low-resource settings	No
Confocal scanning laser ophthalmoscopy	Late 1980s	Germany	Webb et al[35], 1987	Provides high-resolution, high-contrast images; improves diagnostic accuracy for subtle abnormalities	High cost, requires specialized training, limited adoption, particularly in low-resource settings	No
Multispectral Imaging	2012	Canada	Ma et al[36], 2023	Enhances contrast and detail in retinal images by capturing muliple wavelengths of light	High cost, limited availability, not widely adopted in low-resource settings	No
Smartphone-based retinal imaging	Early 2010s	United Kingdom	Kim et al[37], 2018	Cost-effective, portable, accessible; useful in remote and low-resource settings	Variable image quality depending on lighting and operator skill; requires adequate training	No
Hyperspectral imaging	Early 2010s	Canada	Akbari and Kosugi[39], 2009	Captures detailed biochemical information; high accuracy in tissue composition analysis; valuable for early detection	Complex, expensive, not widely available, limited adoption in clinical practice	No
Photoacoustic imaging	Early 2010s	United States	Hu and Wang[43], 2010	Combines laser-induced ultrasound with optical imaging; provides functional assessment of the retina	Still in research phase, high cost, complex, limited clinical application	No
Teleophthalmology	Early 2000s	United States	Whited[44], 2006	Expands access to DR screening, particularly in underserved areas; allows remote retinal imaging and analysis	Dependent on internet connectivity, requires high-quality imaging devices and trained personnel, lack of direct patient interaction	No
AI and ML algorithms	2018, 2020	United States	Esmaeilzadeh[48], 2024	High sensitivity and specificity; automates retinal image analysis; provides immediate diagnostic feedback	High initial investment, requires continuous algorithm updates, data privacy concerns, integration challenges in clinical workflows	Yes

AI: Artificial intelligence; ML: Machine learning.

Table 2 Summary of artificial intelligence/machine learning techniques in diabetic retinopathy detection

Technique	Description	Advantages	Limitations
CNNs	Deep learning model for image analysis; learns hierarchical features	High accuracy, effective for image data	Requires large datasets, computationally intensive
Support vector machines	Supervised learning model; used for classifying pre-extracted features	Robust with small datasets, interpretable results	Less effective with large-scale image data
Random forests	Ensemble learning method using decision trees; used for feature-based classification	Good performance with noisy data	Requires feature extraction, less flexible than CNNs

CNN: Convolutional neural network.

Table 3 Comparison of artificial intelligence/machine learning models and traditional screening methods for diabetic retinopathy

Screening method	Accuracy	Sensitivity	Specificity	Key points
CNNs	High	High	High	Capable of analysing complex retinal images with high accuracy and scalability
Support vector machines	Moderate	Moderate	Moderate	Effective in classifying pre-extracted features but less scalable than CNNs
Random forests	Moderate	Moderate	Moderate	Good for feature extraction-based classification; robust but less flexible
Traditional manual fundus examination	Variable	Low to moderate	Low to moderate	Dependent on the skill of the ophthalmologist; less accessible and scalable

CNN: Convolutional neural network.

Table 4 Artificial intelligence-driven personalized management strategies in diabetic retinopathy

AI application	Ref.	Description	Impact on patient care
Predicting disease progression	Kong and Song[73], 2024	Analyse vast and diverse datasets, including retinal images, genetic information, blood glucose levels, and other patient-specific variables, to identify subtle patterns and predict the likelihood of disease advancement with higher accuracy	Allows for timely intervention and personalized treatment plans
Optimizing treatment regimens	Patibandla et al[74], 2024	Analyse patient data to predict the effectiveness of different treatment options, such as laser therapy or anti-VEGF injections, and recommend the most suitable approach for each individual	Ensures patients receive the most effective treatments based on individual data
Personalizing follow-up schedules	Silva et al[75], 2024	Determine the optimal frequency of eye exams and other monitoring measures, ensuring timely detection of any changes in a patient's condition	Helps in timely detection of changes in the patient's condition

AI: Artificial intelligence.

Table 5 Challenges and solutions for artificial intelligence/machine learning implementation in diabetic retinopathy screening

Challenge	Ref.	Description	Potential solution
Data standardization and interoperability	Mandl et al[76], 2024	Difficulty in integrating AI tools with diverse electronic health record systems	Develop universal data standards and use fast healthcare interoperability resources APIs
		Clinicians need AI tools that seamlessly integrate into their existing workflows without adding complexity or disrupting patient care
		Ensuring scalability, security, and ongoing technical support are critical considerations
Ethical and regulatory concerns	Goldberg et al[77], 2024	Issues related to data privacy, algorithmic bias, and lack of clear regulatory guidelines	Promote diverse datasets, establish clear regulatory frameworks, and ensure data security
Scalability and maintenance	Marvasti et al[78], 2024	Challenges in deploying AI systems across large healthcare networks	Use cloud-based platforms for scalability and provide ongoing technical support

AI: Artificial intelligence; APIs: Application programming interfaces.