As a Research Assistant to Dr. Davis at Penn’s Center for Neuroengineering and Therapeutics, I trained a deep learning model to automate resection cavity segmentation on postoperative MRI of epilepsy patients to help physicians quantify removed brain structures. 

The purpose of the project was to expand on a previous paper about Deep Learning-Based Automated Segmentation of Resection Cavities on Postsurgical Epilepsy MRI by adding a step in the data preprocessing and then training of the model.

A step of data augmentation was added to the pre-processing. Initially, the MRI was transformed into png slices in the axial view. The step added was including sagittal and coronal views in the model’s training. These images were later used to train the model. Additionally, all of the 3D volumetric images were normalized to a standard intensity range 0 to 1.The models have a U-Net CNN architecture and the training was done using the Keras API with TensorFlow backend.

Lastly, a majority vote algorithm was then used to run inference and determine if a voxel should be in the generated mask or not.

After training the 3 different models, running inference, and the majority vote algorithm, we determined that Axial/Coronal model is the highest performer.

Future work includes investigating the sagittal model and see how we can improve its performance.