A Fully Automated Deep Learning Network for Brain Tumor Segmentation

Abstract

We developed a fully automated method for brain tumor segmentation using deep learning; 285 brain tumor cases with multiparametric magnetic resonance images from the BraTS2018 data set were used. We designed 3 separate 3D-Dense-UNets to simplify the complex multiclass segmentation problem into individual binary-segmentation problems for each subcomponent. We implemented a 3-fold cross-validation to generalize the network's performance. The mean cross-validation Dice-scores for whole tumor (WT), tumor core (TC), and enhancing tumor (ET) segmentations were 0.92, 0.84, and 0.80, respectively. We then retrained the individual binary-segmentation networks using 265 of the 285 cases, with 20 cases held-out for testing. We also tested the network on 46 cases from the BraTS2017 validation data set, 66 cases from the BraTS2018 validation data set, and 52 cases from an independent clinical data set. The average Dice-scores for WT, TC, and ET were 0.90, 0.84, and 0.80, respectively, on the 20 held-out testing cases. The average Dice-scores for WT, TC, and ET on the BraTS2017 validation data set, the BraTS2018 validation data set, and the clinical data set were as follows: 0.90, 0.80, and 0.78; 0.90, 0.82, and 0.80; and 0.85, 0.80, and 0.77, respectively. A fully automated deep learning method was developed to segment brain tumors into their subcomponents, which achieved high prediction accuracy on the BraTS data set and on the independent clinical data set. This method is promising for implementation into a clinical workflow.

Keywords: BraTS; Brain tumor segmentation; CNN (convolutional neural networks); Dense UNet; MRI; deep learning; machine learning.

Figure 1.

Schematic representation of the developed algorithm. The input images included T1, T2, T2-FLAIR, and T1 post-contrast (T1C). Whole tumor (WT-net) segments the WT, tumor core (TC-net) segments the TC, and enhancing tumor (EN-net) segments the ET. The output segmented volumes from each of these networks are combined using a triple volume fusion to generate a multiclass segmentation volume.

Figure 2.

Schematic of the Dense UNet Architecture. Each network consisted of 7 dense blocks, 3 transition down blocks, and 3 transition up blocks (A). Each dense block was made of 5 layers connected to each other, with every layer having 4 sublayers connected sequentially (B). The transition-down block consisted of 5 layers connected sequentially (C). The transition-up block was a sequential connection of 4 layers (D).

Figure 3.

Example Segmentation results. High-grade glioma (HGG) and (A) low-grade glioma (LGG) (B). (a) A 2D slice of a postcontrast image, (b) ground truth image, (c) network output without multivolume fusion (MVF), and (d) network output with MVF. The arrows in (c) represent false positives that are successfully eliminated after MVF (d). Color Codes: red = ET, blue = TC (ET + NEN), green = ED; whole tumor = green + blue + red.