Computer Vision
As part of the research requirements for a Computer Vision course I took in the fall semester of 2023, I conducted a project on medical image segmentation using the DRINet model. The goal of this project was to recreate and apply a deep learning model for segmenting medical images from the MICCAI BraTS 2018/2019 datasets, specifically focused on glioma brain tumors. I selected the paper "DRINet for Medical Image Segmentation" by Chen et al. (2018), which proposed a model designed for accurate segmentation in complex medical imaging scenarios, such as brain tumors.
In my project, I recreated the DRINet model and applied it to the BraTS dataset, which consists of MRI scans in four modalities: T1-weighted, T1 post-contrast (T1C), T2-weighted, and FLAIR. Each of these modalities provided important insights into different aspects of the gliomas, including the tumor core, peritumoral edema, and the enhancing tumor. The dataset contained 285 samples of both high-grade (HGG) and low-grade gliomas (LGG), and my task involved segmenting these tumors into the relevant regions.
The first step was preprocessing the images, which involved image normalization and cropping to isolate the tumor region using the SimpleITK library. This allowed me to reduce the dataset to the most relevant areas for analysis, keeping the dataset size at 285 samples.
To improve the performance of the model and reduce overfitting, I applied data augmentation techniques and tuned various hyperparameters. Overfitting was a significant issue during the initial training stages, but through tripling the dataset via augmentation and applying zero-mean, unit-deviation normalization, I was able to mitigate this problem. Due to limited computation, I was not able to train the model using as many epochs as needed.
Reference:
-
Chen, L., Bentley, P., Mori, K., Misawa, K., Fujiwara, M., Rueckert, D., DRINet for Medical Image Segmentation, IEEE Transactions on Medical Imaging, Vol.37, pp.2453-2462, 2018. (10 pages)