A Prior Knowledge Based Tumor and Tumoral Subregion Segmentation Tool for Pediatric Brain Tumors

The WM and GM probabilities are initialized as ICBM atlas priors (affine transformed into the native space) using Advanced Normalization Tools (ANTs [2]). Due to abnormal distributions of CSF (e.g., hydrocephalus) common in brain tumor patients, CSF probability is not initialized as the ICBM atlas prior. Our model assumes CSF to be darkest component/peak among all tissue types under consideration in FLAIR. Figure 1A shows an example histogram from a representative patient with ATRT. The left most peak is CSF and the right most peak is a mixture of WM and GM. The red dotted line is used as the threshold ($th$) for CSF prior estimation, which is determined by the local minima on the right side of the CSF peak. Voxels of intensity value lower than $th$ are initialized as $\Pr(\mathrm{CSF})=0.9$.

Probability of the “other” class is initialized as uniform distribution ($\Pr(\mathrm{other})=0.5$) for non-CSF region.

For each input scan $I$, the probability density function of each class $Y$, denoted as $p_{I}(x|Y)$, is estimated via kernel density estimation using voxels sampled according to prior probability. Because the WM and GM samples can have contaminants (due to inaccurate priors), an outlier detection approach is used to detect and remove the contaminants (key idea proposed in  [7]). Outlier detection is performed by using the Minimum Covariance Determinant estimator, setting the support fraction to 0.5. The posterior probabilities are then calculated for each input according to Bayes’ theorem:

where $\Pr(Y)$ is the prior probability, and $p_{I}(x)=\sum_{Y\in\mathbb{Y}}p_{I}(x|Y)\Pr(Y).$ The averaged posteriors from all inputs, are used as priors in the next iteration, and the probabilities are updated through a total of 3 iterations in our implementation. The final WM and GM masks are generated by thresholding the smoothed posteriors at 0.5.

DeepNeuro is a DL-based open-source toolbox for neuroimaging [3]. It provides a module developed for glioblastoma (GBM) edema and enhancing tumor segmentation. This module takes T1, T1-post, and FLAIR as inputs, and produces a WT mask and enhancing tumor mask. The GBM model has a sequential 3D U-Nets structure, which predicts WT mask first and then uses the predicted WT mask as the fourth input to predict enhancing mask [4]. This model has been trained on both public and private datasets of post-operative GBMs. Because this GBM model has a sequential structure, we were able to modify the implementation to use only the second U-Net to predict enhancing tumor from a given WT mask, for evaluation of enhancing prediction.

Wang et. al. [8] developed an an open-source model that won the 2nd place of MICCAI 2017 BraTS Challenge. The BraTS datasets contain pre-operative multimodal MRI scans of GMB and LGG. The segmentation task is to predict WT, edema, and enhancing tumor given T1, T1-post, T2 and FLAIR. The model consists of three convolutional nerual networks (CNNs) in sequential order. The first CNN (WNet) predicts WT mask, the second CNN (TNet) predicts tumor core (both enhancing and non enhancing tumor) mask within the bounding box defined by the predicted WT mask from the first CNN, then the third CNN (ENet) predicts the enhancing tumor within the bounding box defined by the predicted tumor core mask. Specifically, these CNNs are anisotropic, i.e., each net was trained in axial, sagittal and coronal views respectively using 2D receptive field (a large receptive field in 2D and a relatively small receptive field in the out-plane direction). The authors provided two pretrained models: model15 and model17, trained on BraTS15 and BraTS17, respectively. In addition to these two models, we trained the WNet on our DIPG data (130 cases with manual WT labels) initialized with model17. We label this model as cascaded-CNN-DIPG and add include this method in our comparison for WT segmentation task only. The 4 DIPG cases used for evaluation were not included in the training set for this model. Additionally, we modified the implementation to accept manually labeled WT mask as input and generate subregion masks using TNet and ENet, to evaluate performance of subregion prediction only.

NVIDIA Clara is an application framework optimized for healthcare and life sciences developers. Clara provides a brain tumor segmentation model based on autoencoder regularization (clara_mri_seg_brain_tumors_br16_full_amp) [6], which won the 1st place in the BraTS 2018 challenge. The Clara model takes T1, T1-post, T2, and FLAIR as inputs, and outputs WT mask, tumor core mask, and enhancing mask. Unlike other methods under comparison, this model generates WT mask and subregion masks simultaneously and requires GPU to make predictions.

The WT segmentation results (dice scores) are shown Table 2. Our proposed approach outperformed all other DL-based methods on the LGG and ATRT dataset, because the test data is very different from the training data of these DL-based models. For ATRT in particular, our approach had a mean dice score of 0.753, where other methods were below 0.3. All methods consistently performed better on the DIPG tumors compared to the other two data sets. Regarding the DIPG dataset, the best performer was the Cascaded CNN (DIPG) with a mean dice score of 0.924. This was expected because this model has been trained on a sufficient amount of existing DIPG data (130 cases). Our approach performed nearly identical to the 2nd place model, Cascaded CNN (BraTS17), with dice scores of 0.869 and 0.870 respectively, and is substantially better than the remaining models (0.56, 0.801, and 0.594).

Among all subregions under consideration, we were only able to compare our proposed model with others on enhancing tumor and edema segmentations. None of the existing methods used in comparison are capable of making a prediction regarding necrosis, hemorrhage, CSF and cyst, hence we only compared our results with the manual labels. There results of subregion segmentation are shown in Table 3. Our proposed approach had the highest mean dice score on both enhancing and edema segmentation. The results of other subregions (necrosis, hemorrhage, CSF, and cyst) were reasonable and performance varied across specific subregion. Representative examples of subregion segmentations by our approach are shown in Fig.  3