AUTOMATED DETECTION OF MALARIAL RETINOPATHY USING TRANSFER LEARNING

Area of Science:

• Ophthalmology
• Infectious Diseases
• Artificial Intelligence

Background:

• Cerebral malaria (CM) is a severe neurological complication of malaria, primarily affecting children.
• Malarial retinopathy (MR) presents as specific retinal lesions indicative of CM.
• Current CM diagnosis faces challenges, with up to 23% of patients misdiagnosed due to similar symptoms from other infections, leading to delayed or incorrect treatment.

Purpose of the Study:

• To develop and validate a diagnostic technique utilizing transfer learning for accurate identification of malarial retinopathy (MR).
• To specifically detect retinal hemorrhages and whitening lesions characteristic of MR using retinal images.
• To reduce false positive diagnoses in CM cases by providing a highly specific diagnostic tool.

Main Methods:

• A transfer learning technique was employed using retinal images from three distinct retinal cameras.
• The model was trained to identify key indicators of MR, including hemorrhages and whitening lesions.
• Performance was evaluated based on specificity, sensitivity, and Area Under the Curve (AUC).

Main Results:

• The developed MR detection model achieved 100% specificity and 90% sensitivity.
• The model's performance was quantified with an Area Under the Curve (AUC) of 0.98.
• The algorithm demonstrated high accuracy in identifying MR, even with images from low-cost retinal cameras.

Conclusions:

• The study presents a promising AI-driven approach for the accurate diagnosis of malarial retinopathy (MR).
• This technique offers a highly specific diagnostic tool, crucial for differentiating CM from other conditions and reducing misdiagnosis.
• The potential for using low-cost retinal cameras makes this approach scalable and accessible for improved malaria diagnostics in affected regions.