Brain tumor segmentation with U-Net and FC-DenseNets

Semantic image segmentation predicts class labels for each pixel in the image. In medicine, semantic segmentation is widely used for medical image diagnostics, where machines can augment analysis performed by radiologists with efficiency and scalability. One active area of research is brain tumor segmentation, where deep neural networks are trained to predict where each pixel falls into tumorous tissue classes that were labeled by neuroradiologists. In this work, we used a data set comprising clinically-acquired pre-operative multimodal MRI scans of glioblastoma and lower grade glioma, common types of brain tumors. This data set was preprocessed and labeled by neuroradiologists for BraTS 2020 competiton. We trained two state-of-the-art deep neural network models for image segmentation, U-net and FC-DeepNets and evaluated their performance.

Motivation

We build CNN-based models for pixel-by-pixel segmentation of brain tumor subregions [0: Not tumor, 1: Necrotic/Core, 2: Edema, 3: Enhancing core]. Two models with U-Net and FC-DenseNets architectures were trained and their performance was evaluated using various metrics (see below). The MRI data are collected/preprocessed by the BraTS community for 2020 BraTS challenge.

Models

U-Net comprises symmetrical downsampling, upsampling paths, and skip connections between them (Ronneberger et al., 2015).

Fully-convolutional DenseNets also comprises downsampling (contracting) and upsampling (expanding) paths as well as the skip connections from the downsampling to the upsampling path. The downsampling path incorporated three dense blocks, in which all feature outputs are iteratively concatenated in a feedforward fashion to facilitate the feature reuse (Jegou et al., 2016).

Model evaluation

To assess model performance, we used the held out test set and visualized model predictions and ground truths in tandem for each class and observed that they largely overlapped.

We then quantified model performance using diverse metrics, and compared performance of the two models based on U-Net versus FC-DenseNets. With the current model configurations and dataset, the U-Net outperformed the more complex and deeper FC-DenseNets at least on the versions of the models we configured. Admittedly, there are many model parametric, hyper-parametric, and schematic (e.g. postprocessing) spaces we have not exhaustively explored that might improve model performance, which would be an important direction of future studies.

Documentation

📄 Please find the final report of this project here: Final report
💻 Please find the final presentation of this project here: Final presentation