Retinal Pathology Diagnostics Application:
Objective:
In a country like India where the number of patients significantly outnumber the number of specialist doctors by several times, automation of diagnostics could support the doctors to facilitate faster treatment to patients. Before any algorithm is deployed on a medical device-width it has to be robustly tested and FDA approved. In this project, I worked on developing diagnostic pipelines for detecting retinal pathology using machine learning algorithms as a cloud based application. This tool facilitated pre-clinical testing of the diagnostic pipelines to provide support for FDA approval.
Contributions:
- Trained and tested a multimodal Convolutional Neural Network using Fundus images and patient metadata to detect Diabetic Retinopathy stage.
- Trained and tested a 3D Convolutional Neural Network using OCT DICOM images and patient metadata to detect Age related Macular Degeneration and Diabetic Macular Edema.
- Collaborated with retina specialists to design experiments, data collection and clinical validation and compliance testing of diagnostic pipelines.
- Led a cross functional team of frontend developers and devops engineers to develop a cloud based application for automated diagnosis of patient data for retinal diseases.
Result:
- Achieved an AUC of 0.82 in detecting Age related Macular Degeneration and Diabetic Macular Edema.
- Improved model performance of classifying the stage of Diabetic Retinopathy by 16% in terms of macro average AUC metric.
- Facilitated pre-clinical testing of diagnostic pipelines by retina specialists aiding FDA application process.