Application of Deep Learning in Postoperative Recurrence Prediction for Pediatric Low-Grade Gliomas

Background

Pediatric Low-Grade Gliomas (PLGGs) are one of the most common types of brain tumors in children, accounting for 30%-50% of all central nervous system tumors in children. Although the prognosis of PLGGs is relatively favorable, the risk of postoperative recurrence is difficult to accurately predict using traditional clinical, imaging, and genomic factors. The heterogeneity of postoperative recurrence complicates decision-making regarding postoperative management, particularly in terms of adjuvant therapy and imaging surveillance. Therefore, developing a tool capable of accurately predicting postoperative recurrence risk is crucial for optimizing patient management and improving outcomes.

In recent years, Deep Learning (DL) has made significant progress in medical imaging analysis, particularly in tumor segmentation and prognosis prediction. However, due to the rarity and data scarcity of PLGGs, the application of DL in this field still faces challenges. This study aims to develop a multimodal DL model by combining preoperative magnetic resonance imaging (MRI) and clinical data to improve the prediction of postoperative recurrence risk in PLGGs.

Source of the Paper

This paper is a collaborative effort by a team from multiple institutions, with primary authors including Maryamalsadat Mahootiha, Divyanshu Tak, Zezhong Ye, and others. The corresponding author is Benjamin H. Kann from the Dana-Farber Cancer Institute and Brigham and Women’s Hospital at Harvard Medical School. The paper was published online in advance on August 30, 2024, in the journal Neuro-Oncology, with the DOI 10.1093/neuonc/noae173.

Research Process and Results

1. Dataset and Preprocessing

This study utilized datasets from two institutions: Dana-Farber/Boston Children’s Hospital (DF/BCH) and the Children’s Brain Tumor Network (CBTN). A total of 396 patients were included, with 200 from DF/BCH and 196 from CBTN. All patients underwent surgery and had preoperative T2-weighted MRI scans. Patients with neurofibromatosis were excluded due to their distinct disease trajectory compared to other PLGGs.

MRI images were first converted from DICOM to NIfTI format, and N4 bias field correction was applied to eliminate low-frequency intensity non-uniformity. All scans were resampled to a voxel size of 1×1×3 mm³ and aligned using rigid registration. Brain tissue extraction was then performed using the HD-BET software package.

2. Deep Learning Feature Extraction

A pre-trained 3D tumor segmentation model (based on the UNet architecture) was used to extract imaging features of PLGGs. This model had previously been fine-tuned on CBTN patient data but had not been trained on DF/BCH patient data. Through transfer learning, the researchers extracted 4096 high-dimensional features from the encoder of the segmentation model, which were believed to capture abstract representations of the tumor.

3. Event-Free Survival (EFS) Prediction

The researchers developed a fully automated DL pipeline that integrated the segmentation model, feature extractor, and a three-layer fully connected neural network to predict postoperative EFS. Three models were trained: 1) a model using only clinical features (e.g., age and resection status); 2) a model using only DL-MRI features; and 3) a multimodal model combining clinical and DL-MRI features.

4. Model Training and Validation

The datasets from the two institutions were merged and randomly split in a stratified manner based on EFS status and data source, with 70% used for model development and 30% for testing. The mean performance of the models was evaluated using 3-fold cross-validation. The study also explored the impact of fine-tuning the models on data from different institutions.

5. Model Performance Evaluation

On the test set, the multimodal model achieved a C-index of 0.85 (95% CI: 0.81-0.93), significantly outperforming the models using only DL-MRI features (C-index: 0.79) and only clinical features (C-index: 0.72). The multimodal model also achieved an AUC of 0.88 for 3-year EFS prediction, surpassing the other two models.

6. Risk Stratification

The multimodal model effectively stratified patients into low-risk and high-risk groups, with a 3-year EFS of 92% for the low-risk group and 31% for the high-risk group (p < 0.0001). In contrast, the models using only DL-MRI features or only clinical features performed less effectively in risk stratification.

Conclusions and Significance

This study is the first to demonstrate the application of multimodal DL in predicting postoperative recurrence in PLGGs. By combining preoperative MRI imaging and clinical data, the DL model significantly improved the accuracy of postoperative recurrence risk prediction and provided strong support for postoperative management decisions. The study also demonstrated the effectiveness of transfer learning on small datasets, offering new insights for future applications in rare diseases.

Research Highlights

1. Multimodal Deep Learning: For the first time, MRI imaging features were combined with clinical data, significantly improving the prediction of postoperative recurrence risk in PLGGs.
2. Application of Transfer Learning: Imaging features were extracted using a pre-trained segmentation model, overcoming the issue of data scarcity.
3. Automated Pipeline: A fully automated DL pipeline was developed, capable of rapidly generating personalized EFS prediction curves with high potential for clinical application.

Future Prospects

Although this study achieved significant results, external validation on larger, multicenter datasets is needed to further improve the model’s generalizability. Additionally, future research could further explore the impact of adjuvant therapy on model performance and investigate how to integrate the model into clinical workflows to optimize patient management.