Accuracy of a free, open source, machine learning model, made using teachable machine in detecting retinitis pigmentosa on fundus images

Background and Aims: Assessing the accuracy and the discrimination ability of a free web-based machine learning model to identify the presence of retinitis pigmentosa. Settings and Design: For the study, 373 ultrawide-field autofluorescence images (150, retinitis pigmentosa; 223, normal) were obtained from the patients. The machine learning tool used is called teachable machine that uses TensorFlow.js, a library for machine learning in JavaScript, to train and run the models in the browser of the user itself. Materials and Methods: Two separate deep learning models were created with both models using 85:15 as the ratio of training vs testing images. The images in both models were randomly selected in one case by the teachable machine neural network itself, and in the other, it was manually selected. Results: The accuracy of the first model was found to be 98.25%. The sensitivity of the model was found to be 95.65%, and the specificity was found to be 100%. The accuracy of the second model was found to be 94.74%. The sensitivity and specificity of the model were found to be 91.30% and 97.06%. Discussion: The study found to show that the machine learning model could distinguish between the retinitis pigmentosa eyes and normal, healthy eyes with a high specificity and sensitivity.