Overview / Usage

Our projects aims to bridge the gap formed by ambiguity and misdiagnosis of diseases by training a model to correctly predict a limited number of diseases and serve as helping tool for Doctors to make better decisions (Reference: CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays Using Deep Learning).
Our model being developed is an ensembled Convolutional Neural Network (CNN) trained on 223,414 X-ray images of 64,540 unique patients varying in age ,gender and diseases they have.
There are 5 classes (4 diseases, and one for "No findings").

Methodology / Approach

Proposed research work has been implemented by training an Ensembled Convolutional Neural Network (CNN) using CheXpert Dataset. Use of ensemble model of Densenet, VGG and Inception v3 has been done to achieve good results for the intended task of detecting possibility of disease in a Chest X-Ray image. Entire code for the proposed research work has been written using Keras framework based on Tensorflow. Use of Google Colaboratory and Intel AI DevCloud has been done for training of the model for the intended task. All of the above mentioned pre-trained models are already capable of identifying 1000 different categories of objects due to their intensely dense neural architecture and this capability of the models has been leveraged using the process of fine-tuning the models. In the process of fine tuning few additional layers have been added at the end of the pre trained model. At first the entire layers of the pre trained model are freezed and the layers that were added are trained over Chest X-Ray dataset. Once the additional layers get trained then the entire layers of the model are unfreezed and the training of whole model is done wherein all the layers adjust their weights using the process of Backpropagation in neural networks. Each pre trained model gives a prediction vector as an output. A weighted average of these predictions is taken to estimate the most probable disease present in the X-ray image.