Group Member 1: Cam Backes ([email protected])
Group Member 2: Jon Deaton ([email protected])
Noninvasive methods of brain imaging, most commonly Magnetic Resonance Imaging (MRI), are routinely used to identify and locate tumors in the brain. Currently, brain tumor image segmentation is a time consuming practice which must be performed manually by medical professionals. As such, with the recent emergence of effective computer vision methods, notably convolutional neural networks (CNNs), there is significant practical value in using these tools to automate and improve the accuracy of segmentation. We propose using capsule networks to perform segmentation of brain tumors in MR images.
The Brain Tumor Segmentation (BraTS) dataset was collected by medical professionals from numerous institutions including UPenn’s Center for Biomedical Image Computing and Analysis (CBICA), and consists of 3D MRI brain scans from 333 individuals with brain tumors, along with age and survival information for each individual, and tumor segmentation labels for tumor pixels manually-revised by expert board-certified neuroradiologists.
We have chosen to apply Capsule Network techniques to this problem because of recent advances in training Capsule Networks that have shown to improve upon shortcomings of CNNs. The most notable difficulty associated with this project is the small size of our dataset. CNNs typically require thousands of images for training, so we will need to be particularly cautious not to overfit our model to the training data. However, a primary source of our inspiration to use Capsule Networks is their demonstrated ability to use training data more efficiently, thus requiring fewer training examples than CNNs. Another challenge comes from the fact that MRI scans are three dimensional.
Because this dataset is made publicly accessible for the purpose of promoting growth in the field of biomedical imaging analysis, there are numerous relevant papers that elucidate the nature of the dataset and explore the use of CNNs and other models for image segmentation. We’ve listed a few of these papers in the references section below.
We will evaluate model performance using the metrics defined in Mense et al., which are used to evaluate models submitted to the MICCAI BraTS challenge, an annual competition in which researchers’ algorithms are judged based on tumor segmentation performance. For each scan, the Dice, sensitivity, and specificity metrics are computed based on discrepancies between predicted tumor segmentation, and the true segmentation by doctors. We will evaluate our model by computing these metrics using our validation set and comparing our results with those obtained by previous challenge participants.
We will display images of brains with our segmentation predictions overlaid, as well as the real tumor segmentations for comparison. We will also show histograms of our algorithm’s Dice, sensitivity, and specificity scores on the test set versus those of past competition entries, as well descriptive statistics that further detail our model’s performance.
Dataset overview: The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS)
Successful deep learning approaches: Automatic Brain Tumor Segmentation using Cascaded Anisotropic Convolutional Neural Networks Brain Tumor Segmentation with Deep Neural Networks