top of page

Thank you for your submission!

Utilization of automated deep learning approach toward detection of ocular toxoplasmosis using fundus photographs

Muhammad Hassan


Hassan, Muhammad; Ormaechea, Maria S; Halim, Muhammad S; Uludag, Gunay; Schlaen, Ariel; Kesim, Cem ; Colombero, Daniel; Ozdemir, Huseyin; Rudzinski, Marcelo; Subramaniam, Mahadevan; Ozdal, Pinar; Chundi, Parvathi; Nguyen, Quan D; Hasanreisoglu, Murat



1. Stanford University Department of Ophthalmology, Palo Alto, CA, United States.

2. Hospital Universitario Austral, Pilar, Argentina.

3. Universidad Nacional de Rosario, Rosario, Argentina.

4. Koc Universitesi Tip Fakultesi, Ankara, Turkey.

5. University of Nebraska Omaha, Omaha, NE, United States.

6. SBU Ulucanlar Goz Egitim Ve Arastirma Hastanesi, Ankara, Ankara, Turkey.

Purpose: Ocular Toxoplasmosis is often a clinical challenge that may require an expert opinion. A correct and timely diagnosis of active disease is the key to preventing significant vision loss from the disease. In this study, we aim to develop a deep learning algorithm without coding for the differentiation of Ocular Toxoplasmosis fundus images from a normal fundus image.


Methods: Patients with a confirmed diagnosis of Ocular Toxoplasmosis who had fundus photos available for the diseased eye/s were included in the study. Patients were excluded if they had other concomitant ocular diseases, had ungradable imaging data, or if the diagnosis was uncertain. The Ocular Toxoplasmosis fundus photos were obtained from uveitis centers in Argentina, Turkey, and USA. The healthy fundus photos were obtained from publicly available databases. The fundus photos included in the study were a combination of standard fundus photos and ultrawide field fundus photos. A deep learning model using the automated machine learning (AutoML) vision platform from Google LLC (Menlo Park, CA) was trained using 441 Ocular Toxoplasmosis and 103 normal images followed by validation using 54 Ocular Toxoplasmosis and 12 normal images. The model was then tested using 57 Ocular Toxoplasmosis and 15 normal images. The area under the precision-recall curve (AUPRC) was plotted and sensitivity, specificity, positive predictive value (PPV), and accuracy (AC) were calculated.


Results: A total of 552 Ocular Toxoplasmosis patient images were compared to 130 healthy images. AUPRC for the dataset was found to be 0.99 (Figure 1A). The sensitivity, specificity, PPV, and AC of the model were 96.5%, 100%, 100%, and 97%. Figure 1B also shows the confusion matrix of the model.


Conclusion: Clinician-derived automated machine learning model developed without coding was able to differentiate Ocular Toxoplasmosis from normal images. This model has the potential to be developed further to aid physicians in the diagnosis of Ocular Toxoplasmosis. Additionally, AutoML can enable clinician-derived discovery of disease biomarkers.


bottom of page