Deep learning strategies for semantic segmentation of pediatric brain tumors in multiparametric MRI

Abstract

Automated segmentation of pediatric brain tumors (PBTs) can support precise diagnosis and treatment monitoring, but it is still poorly investigated in literature. This study proposes two different Deep Learning approaches for semantic segmentation of tumor regions in PBTs from MRI scans. Two pipelines were developed for segmenting enhanced tumor (ET), tumor core (TC), and whole tumor (WT) in pediatric gliomas from the BraTS-PEDs 2024 dataset. First, a pre-trained SegResNet model was retrained with a transfer learning approach and tested on the pediatric cohort. Then, two novel multi-encoder architectures leveraging the attention mechanism were designed and trained from scratch. To enhance the performance on ET regions, an ensemble paradigm and post-processing techniques were implemented. Overall, the 3-encoder model achieved the best performance in terms of Dice Score on TC and WT when trained with Dice Loss and on ET when trained with Generalized Dice Focal Loss. SegResNet showed higher recall on TC and WT, and higher precision on ET. After post-processing, we reached Dice Scores of 0.843, 0.869, 0.757 with the pre-trained model and 0.852, 0.876, 0.764 with the ensemble model for TC, WT and ET, respectively. Both strategies yielded state-of-the-art performances, although the ensemble demonstrated significantly superior results. Segmentation of the ET region was improved after post-processing, which increased test metrics while maintaining the integrity of the data.

Keywords: Deep learning; MRI; Pediatric brain tumor; Tumor segmentation.

Fig. 1

Redefinition of BraTS-PEDs 2024 annotations for the segmentation task.

Fig. 2

Pre-processing pipeline of the volumetric data.

Fig. 3

3-encoder architecture model.

Fig. 4

Residual blocks and SE blocks explanation. (a) Residual blocks consist of [3, 3, 3] 3D convolutions with stride 1 and padding 1 to preserve spatial dimensions while reducing channel counts. The convolutional outputs are followed by Instance Normalization and LeakyReLU activation, introducing non-linearity. The input is then summed with the output to implement residual connections, ensuring efficient gradient flow and retaining original input features. (b) The SE block employs adaptive average pooling during the squeeze phase, reducing spatial dimensions and retaining only per-channel information. In the excitation phase, fully connected layers compress and restore channel dimensions, incorporating Gaussian Error Linear Unit (GeLU) activation and dropout (20%) to prevent overfitting. The block outputs channel-wise weights, which modulate the input tensor to emphasize important channels.

Fig. 5

4-encoder architecture model.

Fig. 6

Variation of the ET Dice Score on the validation set across with the ET/WT thresholds tested for the two models.

Fig. 7

Slice 22 of the ground truth mask alongside the predicted masks from the ensemble model and SegResNet for patient 00239-000 in the test set.

Fig. 8

Slice 62 of the ground truth mask alongside the predicted masks of ET region from the ensemble model and SegResNet for patient 00028-000 in the test set.