A brain MRI dataset and baseline evaluations for tumor recurrence prediction after Gamma Knife radiotherapy

Yibin Wang, William N Duggar, David Michael Caballero, Toms Vengaloor Thomas, Neha Adari, Eswara Kumar Mundra, Haifeng WangScientific 2023 · DOI

Abstract

Prediction and identification of tumor recurrence are critical for brain cancer treatment design and planning. Stereotactic radiation therapy delivered with Gamma Knife has been developed as one of the common treatment approaches combined with others by delivering radiation that targets accurately on the tumor while not affecting nearby healthy tissues. In this paper, we release a fully publicly available brain cancer MRI dataset and the companion Gamma Knife treatment planning and follow-up data for the purpose of tumor recurrence prediction. The dataset contains original patient MRI images, radiation therapy data, and clinical information. Lesion annotations are provided, and inclusive preprocessing steps have been specified to simplify the usage of this dataset. A baseline framework based on a convolutional neural network is proposed companionably with basic evaluations. The release of this dataset will contribute to the future development of automated brain tumor recurrence prediction algorithms and promote the clinical implementations associated with the computer vision field.

Introduction

Distinguishing between radionecrosis (radiation-induced tissue death) and actual tumor recurrence is a significant clinical challenge following Gamma Knife stereotactic radiosurgery. While Gamma Knife allows for precise, high-dose radiation targeting, the resulting post-treatment scans often present ambiguous features. Accurate prediction of recurrence is vital for timely intervention and treatment planning. However, existing AI models often lack the multi-modal context—specifically the spatial distribution of radiation dose—necessary to make this distinction effectively.

The Multi-Modal Dataset

To address this gap, we present a first-of-its-kind public dataset hosted on The Cancer Imaging Archive (TCIA). This dataset integrates structural brain MRIs with 3D radiation dose maps (RTDose) and clinical metadata.

Overview of the multi-modal dataset components, including MRI with lesion annotations, spatial Dose Info, and categorical clinical data.

The cohort consists of 47 patients from the University of Mississippi Medical Center, encompassing 244 annotated lesions. A key characteristic of this dataset is the clinical imbalance: 221 lesions remained stable, while 23 showed recurrence. Data is provided in raw DICOM, NRRD, and preprocessed numpy formats to facilitate immediate research use.

Methodology

The proposed pipeline focuses on bridging the gap between radiology and radiation oncology data through three main steps:

  1. Multi-Modal Integration: The model utilizes structural MRI (T1 MPRAGE with contrast), 3D RTDose maps, and clinical data (age, gender, primary diagnosis).
  2. Spatial Alignment: To ensure voxel-wise correspondence, affine transformations were employed to resample the dose data into the MRI coordinate space.
  3. Baseline 3D CNN: A multi-input 3D Convolutional Neural Network was developed. Parallel 3D convolutional layers process the MRI and Dose volumes, which are then concatenated with clinical feature vectors for final classification.
The spatial relationship between the structural MRI and the corresponding radiation dose distribution is a critical feature for recurrence prediction.

Results and Discussion

Baseline evaluations highlight the inherent difficulty of predicting recurrence in a highly imbalanced clinical setting. While the model achieved a high overall accuracy of 90.1% and a specificity of 89.0%, its sensitivity was notably low at 10.0%.

The "Dose Matters" hypothesis suggests that recurrence often occurs at the boundaries of high-dose zones. By providing these maps, we enable models to learn spatial relationships that standard MRIs alone cannot capture. However, the rarity of recurrence events (~10%) remains a significant hurdle, leading the baseline model to favor the majority "stable" class.

Conclusion and Future Work

This dataset provides the community with the necessary tools to tackle the complex problem of post-radiosurgery recurrence prediction. Future research should focus on addressing class imbalance through specialized loss functions or synthetic data generation, as well as improving model interpretability to aid clinical decision-making.

Dataset available at: https://doi.org/10.7937/xb6d-py67

Summary updated

Abstract

Prediction and identification of tumor recurrence are critical for brain cancer treatment design and planning. Stereotactic radiation therapy delivered with Gamma Knife has been developed as one of the common treatment approaches combined with others by delivering radiation that targets accurately on the tumor while not affecting nearby healthy tissues. In this paper, we release a fully publicly available brain cancer MRI dataset and the companion Gamma Knife treatment planning and follow-up data for the purpose of tumor recurrence prediction. The dataset contains original patient MRI images, radiation therapy data, and clinical information. Lesion annotations are provided, and inclusive preprocessing steps have been specified to simplify the usage of this dataset. A baseline framework based on a convolutional neural network is proposed companionably with basic evaluations. The release of this dataset will contribute to the future development of automated brain tumor recurrence prediction algorithms and promote the clinical implementations associated with the computer vision field. The dataset is made publicly available on The Cancer Imaging Archive (TCIA) ( https://doi.org/10.7937/xb6d-py67 ).
BibTeX
@inproceedings{wang2023brain,
  author = {Yibin Wang and William N Duggar and David Michael Caballero and Toms Vengaloor Thomas and Neha Adari and Eswara Kumar Mundra and Haifeng Wang},
  title = {{A brain MRI dataset and baseline evaluations for tumor recurrence prediction after Gamma Knife radiotherapy}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41597-023-02683-1},
}
Figure