Deep learning now allows us to easily create artificial intelligence to help automate analysis techniques which were previously thought impossible for computers. In this project, I use deep learning model to accurately diagnose pneumonia through chest x-ray image inputs and UiPath automating the deep learning training and testing process.
1. I chose a simple architecture to work on.
2. Because first few layers capture general details like colour blobs, patches, edges, so, instead of randomly initialize weights for these layers, it would be much better if you fine-tune them by importing the pre-trained weights from imagenet.
3. I added layers that introduce a lesser number of parameters. For example, SeparableConv in Keras is a good replacement for Conv layer. It introduces less number of parameters and filters comparing to normal convolution while capturing more information.
4. I also added a batch norm with convolutions. For a deep network, the batch norm is an efficient choice.
5. I put dense layers at the end and trained with a higher learning rate and experiment with the number of neurons in the dense layers. Once the model learnt a good depth, I started training my network with a lower learning rate along with decay.
False Negative: 42%;
After saving the best deep learning model, I built a workflow in UiPath to automate my testing process. The workflow will load the deep learning model, ask the users to select their chest Xray image and use the pre-trained model to test on the selected image. The diagnosis results will be displayed in a message box along with the selected image. All results will be saved into a log so that users can keep track of their testing process.