Quality control of brain tumor segmentations

Publication: QCResUNet: Joint Subject-Level and Voxel-Level Prediction of Segmentation Quality

This project looks into building a prototype deep learning model to grade computationally-generated brain tumor segmentations in the absence of ground truth annotations. The project is motivated by uncertainty estimation approaches in medical imaging AI when making predictions on unseen data. We propose building a secondary deep learning model that takes as input a brain tumor segmentation, generated by pretrained models like nnUNet, and predicts a Dice score and a pixel-wise error map of the segmentation.

My contributions to the project include generating a pipeline for generating grouth-truth-predicted 2D brain tumor segmentation pairs from 3D volumetric brain MRI scans, building the prototype of the model, addressing challenges in predicting Dice scores for slices on the border of brain tumors. I also contribute to literature review and discussions for alternative architectures and training objectives that lead to the final design of the model.

I’d like to thank Dr. Aristeidis Sotiras for the opportunity to work on this project!