OMT and tensor SVD–based deep learning model for segmentation and predicting genetic markers of glioma: A multicenter study

Zhengyang Zhu, Han Wang, Tiexiang Li, Tsung-Ming Huang, Huiquan Yang, Zhennan Tao, Zhong-Heng Tan, Jianan Zhou, Sixuan Chen, Meiping Ye, Zhiqiang Zhang, Feng Li, Dongming Liu, Maoxue Wang, Jiaming Lu, Wen Zhang, Wenlin Zhou, Qian Chen, Zhuoru Jiang, Futao Chen, Xin Zhang, Wen‐Wei Lin, Shing‐Tung Yau, Bing ZhangProceedings of the National Academy of Sciences 2025 · DOI

TL;DR

This paper introduces OMT-APC, a novel deep learning model that uses Optimal Mass Transport (OMT) and tensor Singular Value Decomposition (SVD) to segment brain tumors and predict their genetic markers (WHO grade, IDH mutation, 1p/19q codeletion) from MRI scans. This approach achieves high accuracy across diverse datasets, outperforming human radiologists and offering a non-invasive method for crucial diagnostic information.

What Problem Does This Paper Address?

Accurate genetic profiling of gliomas is critical for determining prognosis and guiding treatment. Traditionally, this requires invasive tissue biopsies, which are not always feasible or safe. While MRI can reveal tumor characteristics, standard deep learning models often struggle with the variability across different imaging centers and protocols (multicenter datasets). Furthermore, predicting multiple genetic markers simultaneously (multitask learning) can be challenging. This paper aims to bridge this gap by developing a robust, non-invasive method for both segmenting gliomas and predicting key genetic markers using preoperative MRI data from a large, diverse, multicenter cohort.

Methods

The core of the OMT-APC model lies in two novel components:

  • Optimal Mass Transport (OMT) for Data Normalization and Segmentation: OMT is employed to transform irregular MRI brain images into standardized tensors. This process preserves the global structure of the brain while enhancing the tumor region within the tensor. This normalization is crucial for handling variations across different datasets and imaging centers, leading to improved generalization. The OMT-derived tensors are then used as input for an nnU-Net model for automated tumor segmentation.
  • Algebraic Preclassification (APC) using Multimode Tensor SVD: To predict genetic markers, the researchers developed an Algebraic Preclassification (APC) model. This model leverages multimode tensor Singular Value Decomposition (SVD) applied to the OMT tensors. SVD decomposes the tensor data into components that capture essential information, allowing for an efficient and reliable estimation of preclassification probabilities. This algebraic approach is integrated with a deep learning framework (SE-ResNet) to form the OMT-APC model, enabling multitask classification.

Datasets

The study utilized a substantial multicenter dataset comprising 3,565 glioma patients from 16 datasets spanning Asia, Europe, and America.

  • Training and Internal Validation: 2,551 patients from 5 datasets.
  • Independent External Test: 1,014 patients from 11 datasets, including 242 patients from The Cancer Genome Atlas (TCGA).

The study also mentions using the BraTS 2023 Adult Glioma Challenge dataset for evaluating the tumor segmentation performance.

Key Insights and Ideas

  • OMT as a powerful data harmonizer: The OMT technique proves highly effective in normalizing diverse MRI data from multiple centers, a significant hurdle in developing generalizable AI models for medical imaging.
  • Tensor representation for preserving global structure: Transforming 3D MRI into tensors with OMT helps retain crucial global anatomical information, which might be lost in traditional cropping or slicing methods.
  • Algebraic preclassification complements deep learning: The APC model, derived from tensor SVD, provides an efficient way to estimate probabilities that enhance the deep learning model's performance, especially for multitask classification.
  • Multitask learning is feasible and effective: The OMT-APC model demonstrates the ability to simultaneously predict WHO grade, IDH mutation, and 1p/19q codeletion status with high accuracy.
  • Outperforming human experts: The model's superior performance compared to radiologists highlights the potential of AI in complex diagnostic tasks.

Key Results

  • Tumor Segmentation: The OMT-based nnU-Net model achieved a mean lesion-wise Dice score of 0.880 on the BraTS 2023 validation set, comparable to top-performing models.
  • Genetic Marker Prediction (TCGA Test Set):
    • WHO Grade: Accuracy of 0.855, AUC of 0.845.
  • IDH Mutation: Accuracy of 0.917, AUC of 0.908.
  • 1p/19q Codeletion: Accuracy of 0.809, AUC of 0.769.
  • Comparison to Radiologists: The OMT-APC model outperformed four radiologists in all prediction tasks (WHO grade, IDH mutation, and 1p/19q codeletion) on the TCGA dataset.
  • Generalizability: The model showed strong performance across internal validation, external test sets, and the TCGA benchmark, indicating good generalization capabilities.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the OMT method for segmentation can be further improved, particularly in enhancing edge recognition of the Whole Tumor (WT) region.
  • While the model uses T1CE and T2WI images, it doesn't incorporate other advanced MRI modalities like diffusion-weighted or perfusion-weighted imaging, which could potentially offer more information.
  • Future Work:
    • The authors plan to focus on optimizing the OMT density function using machine learning to improve WT segmentation accuracy.
  • They aim to develop more advanced algebraic and geometric techniques to extract feature information from medical imaging for gene classification, diagnosis, and treatment planning.
  • Future research will explore incorporating other medical imaging modalities to further enhance diagnostic capabilities.

Summary updated

Abstract

Glioma is the most common primary malignant brain tumor and preoperative genetic profiling is essential for the management of glioma patients. Our study focused on tumor regions segmentation and predicting the World Health Organization (WHO) grade, isocitrate dehydrogenase (IDH) mutation, and 1p/19q codeletion status using deep learning models on preoperative MRI. To achieve accurate tumor segmentation, we developed an optimal mass transport (OMT) approach to transform irregular MRI brain images into tensors. In addition, we proposed an algebraic preclassification (APC) model utilizing multimode OMT tensor singular value decomposition (SVD) to estimate preclassification probabilities. The fully automated deep learning model named OMT-APC was used for multitask classification. Our study incorporated preoperative brain MRI data from 3,565 glioma patients across 16 datasets spanning Asia, Europe, and America. Among these, 2,551 patients from 5 datasets were used for training and internal validation. In comparison, 1,014 patients from 11 datasets, including 242 patients from The Cancer Genome Atlas (TCGA), were used as independent external test. The OMT segmentation model achieved mean lesion-wise Dice scores of 0.880. The OMT-APC model was evaluated on the TCGA dataset, achieving accuracies of 0.855, 0.917, and 0.809, with AUC scores of 0.845, 0.908, and 0.769 for WHO grade, IDH mutation, and 1p/19q codeletion, respectively, which outperformed the four radiologists in all tasks. These results highlighted the effectiveness of our OMT and tensor SVD-based methods in brain tumor genetic profiling, suggesting promising applications for algebraic and geometric methods in medical image analysis.
BibTeX
@inproceedings{zhu2025omt,
  author = {Zhengyang Zhu and Han Wang and Tiexiang Li and Tsung-Ming Huang and Huiquan Yang and Zhennan Tao and Zhong-Heng Tan and Jianan Zhou and Sixuan Chen and Meiping Ye and Zhiqiang Zhang and Feng Li and Dongming Liu and Maoxue Wang and Jiaming Lu and Wen Zhang and Wenlin Zhou and Qian Chen and Zhuoru Jiang and Futao Chen and Xin Zhang and Wen‐Wei Lin and Shing‐Tung Yau and Bing Zhang},
  title = {{OMT and tensor SVD–based deep learning model for segmentation and predicting genetic markers of glioma: A multicenter study}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2025},
  doi = {10.1073/pnas.2500004122},
}
Figure