Network's outputs on a test image as it undergoes training.
I collaborated with davidguzmanr and roher1727 to implement a U-Net Convolutional Neural Network capable of identifying pneumonia indicators generated by COVID-19 in radiology images. For this task, we used a small annotated dataset that was available in the early stages of the pandemic (April 2020).
We applied transfer learning in order to fine-tune a pretrained MobileNetV2 model. In order to overcome the small dataset, we used elastic deformations and random rotations on the radiology images to augment our dataset.
Example of original and augmented images used.
In order to improve performance, we tried other approaches such as aggregating several networks with mixed results. Our final model gave impressive qualitative results overall, closely matching the predictions by trained medical professionals on an evaluation partition of the dataset.
The repository with the model and code can be found here.
