A deep learning neural network utilizing Keras an Tensorflow to create a model based on Convolutional Neural Networks (CNNs) and Residual blocks to detect the severity of Diabetic Retinopathy in fundus photographs.

https://www.coursera.org/account/accomplishments/certificate/49HVF7FYZE3R

Abstract

Purpose: Assess applicability of machine learning models to preoperative clinical and Scheimpflug tomographic data obtained from Oculus Pentacam HR to identify candidates for photorefractive keratectomy (PRK).

Methods: Retrospective, single center, proof of concept study. Preoperative Pentacam data was gathered from 166 patients (332 eyes). Six machine learning classifiers were trained and validated on 82 feature vectors to predict candidacy for PRK against expert clinical evaluation. Performance was assessed by area under the receiver operator curve (AUC) following 10-fold cross validation.

Results: AUC of the optimized Logistic Regression (LR), K Nearest Neighbor (KNN), Multilayer Perception (MLP), Random Forest (RF), Support Vector Classifier (SVC), and AdaBoost models was 0.929, 0.762, 0.842, 0.973, 0.925, and 0.957 respectively. The RF model out performed KNN (p=<0.0001), and MLP (p<0.0001), but was not statistically superior to LR (p=0.088), SVC (p=0.068), or AdaBoost (p=0.469).

Conclusion: RF, LR, SVC, and AdaBoost were accurate and effective in predicting PRK candidacy, while KNN and MLP were the least powerful in this small population.

Created an automated process to detect and classify chest X-ray images using deep learning residual neural networks.

https://www.coursera.org/account/accomplishments/certificate/DJMKCYMMFA93