Improving the Segmentation of Pediatric Low-Grade Gliomas through Multitask Learning

Information Science Artificial Intelligence Pediatrics Oncology Medical Imaging Medicine
multitask learning pediatric low-grade gliomas MRI tumor segmentation deep learning

Improved Segmentation of Pediatric Low-Grade Gliomas Through Multitask Learning

Background Introduction

The segmentation of pediatric brain tumors is a critical task in tumor volume analysis and artificial intelligence algorithms. However, this process is time-consuming and requires the expertise of neuroradiologists. Although significant research has focused on optimizing adult brain tumor segmentation, studies on AI-guided pediatric tumor segmentation are very limited. Moreover, MRI signal characteristics between pediatric and adult brain tumors differ, necessitating segmentation algorithms specifically designed for pediatric brain tumors. Therefore, this paper proposes adding a classifier for genetic changes in brain tumors as an auxiliary task to the main network, enhancing segmentation results through Deep Multitask Learning (DMTL).

Source of the Paper

This study was conducted by researchers Partoo Vafaeikia, Matthias W. Wagner, Cynthia Hawkins, Uri Tabori, Birgit B. Ertl-Wagner, and Farzad Khalvati. The affiliated institutions include The Hospital for Sick Children in Toronto, Canada, University of Toronto, and the Vector Institute. The paper was published at the IEEE Engineering in Medicine & Biology Society (EMBC) 2022 annual international conference.

Workflow

Research Methods and Experimental Setup

The study proposed an MRI-based model for the segmentation of Pediatric Low-Grade Gliomas (PLGGs) while incorporating tumor genetic change classification as an auxiliary task. The specific workflow is as follows:

  1. Data Preparation: Data was collected from FLAIR MRI sequences of 311 pediatric patients treated at the children’s hospital from 2000 to 2018. The tumor ground-truth segmentation was provided by two neuroradiologists, and relevant genetic markers were assessed through biopsy.

  2. Model Architecture: A U-Net-based network was trained on the data. To generate classification outputs, a fully connected layer branch was added to the bottleneck layer of the U-Net, enabling shared learning between the segmentation and classification tasks. The study had received approval from the Children’s Hospital Research Ethics Committee, and the requirement for informed consent was waived by the ethics committee.

  3. Segmentation Task: Using the U-Net network based on an Encoder-Decoder architecture, the encoder part mainly extracted image features, while the decoder part reconstructed the segmentation mask. To ensure consistency in data dimensions and coordinates, all images were registered to the MRI-based SRI24 brain template before executing the segmentation algorithm.

  4. Classification Task: This task identified three major genetic changes in PLGGs: BRAF V600E mutation, BRAF fusion, and other rare molecular markers. The classification network was added to the bottleneck part of the U-Net encoder, enhancing the model’s generalization ability through shared learning.

  5. Loss Function and Optimization: The model was optimized using the Adam optimizer with specific loss functions (Soft Dice Loss for segmentation and weighted cross-entropy loss for classification).

Main Results

Both single-task and multitask models were trained on the same training, validation, and test datasets. The results indicated:

  • Across four tests, the multitask learning method’s performance in the segmentation task improved on average by 3% (from 0.77 to 0.80) in the validation set and 4% (from 0.74 to 0.78) in the test set.
  • In the remaining two tests, the performance of multitask and single-task models was comparable.
  • Overall, the average segmentation performance of the multitask learning method improved by 2.10% in the validation set and 3% in the test set.

This performance improvement is primarily due to the auxiliary classification task, allowing the model to learn useful information related to genetic features, thereby enhancing data representation, reducing overfitting, and improving model generalization.

Data Display

Some experimental results are shown in the table below, demonstrating the performance differences between single-task and multitask models under different data division settings (changing the division of training, validation, and test datasets each time):

| Validation Dice | Test Dice | | Single-task | Multitask | Single-task | Multitask | | – | – | – | – |