# Prediction of brain tumor recurrence location based on multi-modal fusion and nonlinear correlation learning

**Authors:** Tongxue Zhou, Alexandra Noeuveglise, Romain Modzelewski, Fethi Ghazouani, Sébastien Thureau, Maxime Fontanilles, Su Ruan  
**arXiv:** [2304.13725](https://arxiv.org/abs/2304.13725)  
**Date:** 2023-04-11  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2304.13725v1/

## Key points

- Brain tumor is a serious and often fatal disease, with high-grade tumors being particularly prone to recurrence even after standard treatment.
- Predicting the location of brain tumor recurrence is crucial for effective treatment planning and can potentially prolong patient survival time.
- Researchers have developed a deep learning-based brain tumor recurrence location prediction network that uses transfer learning to improve prediction accuracy on small datasets.
- The first step in the process is training a multi-modal brain tumor segmentation network on the public dataset BraTS 2021.
- A multi-scale multi-channel feature fusion model and nonlinear correlation learning module are developed to learn effective features.
- Two decoders are constructed to jointly segment the present brain tumor and predict its future recurrence location.
- This approach represents the first work that can both segment the present tumor while simultaneously predicting future tumor recurrence location, making treatment planning more efficient and precise.
- Experimental results demonstrate the effectiveness of this proposed method in predicting brain tumor recurrence location from limited datasets.

## Summary

Brain tumor is a serious and often fatal disease, with high-grade tumors being particularly prone to recurrence even after standard treatment. Predicting the location of brain tumor recurrence is crucial for effective treatment planning and can potentially prolong patient survival time. To address this issue, researchers have developed a deep learning-based brain tumor recurrence location prediction network that uses transfer learning to improve prediction accuracy on small datasets. The first step in the process is training a multi-modal brain tumor segmentation network on the public dataset BraTS 2021. The pre-trained encoder is then transferred to a private dataset for extracting rich semantic features. A multi-scale multi-channel feature fusion model and nonlinear correlation learning module are developed to learn effective features. The correlation between multi-channel features is modeled by a nonlinear equation, and Kullback-Leibler divergence is used to measure similarity between distributions of original features of one modality and estimated correlated features of another modality. A correlation loss function is designed based on this divergence to maximize similarity between the two feature distributions. Finally, two decoders are constructed to jointly segment the present brain tumor and predict its future recurrence location. This approach represents the first work that can both segment the present tumor while simultaneously predicting future tumor recurrence location, making treatment planning more efficient and precise. Experimental results demonstrate the effectiveness of this proposed method in predicting brain tumor recurrence location from limited datasets. This research provides an important step forward in improving our ability to predict where brain tumors will recur, which could ultimately lead to better outcomes for patients suffering from this devastating disease.