Useful Features for Overall Survival Prediction in Low-Grade Glioma Patients

Academic Background

Glioma is a type of neoplastic growth in the brain that usually poses a serious threat to the patients’ lives. In most cases, glioma eventually leads to the death of the patient. The analysis of glioma typically involves examining pathological slices of brain tissue under a microscope. Although brain tissue pathology images have great potential in predicting patients’ overall survival (OS), due to the unique characteristics of brain tissue pathology, these images are seldom used as the sole predictive factor. Predicting the overall survival rate of early-stage glioma patients using pathological images holds significant value for treatment and quality of life. In this study, the authors explored the possibility of predicting the overall survival rate (OS) of low-grade glioma (LGG) patients using deep learning models in combination with simple descriptive data such as age and glioma subtypes.

Source of Study

This paper was co-authored by Elisa Warner, Xuelu Li, Ganesh Rao, Jason Huse, Jeffrey Traylor, Visweswaran Ravikumar, Vishal Monga, and Arvind Rao. The authors come from the University of Michigan, Amazon, Pennsylvania State University, Baylor College of Medicine, MD Anderson Cancer Center, and the University of Texas Southwestern Medical Center. The paper was published at the 2022 IEEE Engineering in Medicine & Biology Society (EMBC) conference.

Study Content and Methods

a) Research Process

1. Data Acquisition and Preprocessing: The research team obtained 841 pathological slice images from 202 clinically diagnosed glioma patients at the MD Anderson Cancer Center between 1997 and 2015. Patients must have been diagnosed with low-grade glioma (WHO grade 2) at the time of illness and possess pathological images. To adapt to image processing, the images were sliced into 224x224 pixel units, excluding units with more than 40% of blank spaces or pen markings, and underwent staining standardization and artifact removal.

2. Baseline and Comparison Models: The baseline model used an improved VGG16 network, and it was compared to the authors’ custom Regularized Feature Decomposition Deep Network (RFD-Net). These models were used to extract features from the images, converting them into vectorized features through convolutional layers.

3. Ablation Study: To understand the importance of simple clinical data (such as age and glioma subtype) in predicting overall survival, the research team conducted a series of ablation experiments, systematically blocking the types of model inputs and evaluating model performance.

4. Model Evaluation: The models’ performance was evaluated by calculating the slide-level and tile-level AUC (Area Under the ROC Curve) and using cross-validation and a holdout set to ensure the robustness of the models.

b) Main Research Results

1. Model Performance: Through cross-validation and holdout set testing, the RFD-Net model outperformed the baseline model in every instance. The best model combined image and age information, with an AUC of 83.7 for both the test set and the holdout set.

2. Ablation Test Results: The ablation tests showed that in addition to image features, the inclusion of age information improved the model’s robustness and AUC performance. However, including subtype information could have a negative effect.

c) Research Conclusions and Significance

This study demonstrates that using pathological images as the primary data source, combined with basic clinical information, can construct meaningful models for predicting the overall survival rate of low-grade glioma patients. The specific conclusions are as follows:

1. Scientific Value: It shows that pathological images have high potential in predicting the overall survival rate of brain tissues, especially by focusing on distinctive features in the images (such as nuclear features) rather than all parts of the images.

2. Practical Value: Age information is a primary variable for predicting the overall survival rate of brain tissues, which helps improve the model’s robustness and predictive performance. This research result supports the use of age as a prognostic predictor in clinical practice.

d) Study Highlights

1. Important Findings: The RFD-Net model excels in extracting and utilizing distinguishing features in brain pathological images, helping to improve the accuracy of overall survival prediction.

2. Key Issues Addressed: The study addresses the challenge of using brain tissue pathological images for overall survival prediction and proposes a new method focusing on distinguishing features in the images.

3. Innovative Methods: The study employs the Regularized Feature Decomposition Deep Network (RFD-Net), effectively distinguishing between shared and distinguishing features in the images, enhancing predictive capability.

e) Future Development

The research team plans to construct a deep learning model that can reconstruct key image features from distinguishing feature vectors, further increasing the interpretability of the RFD-Net model. Additionally, extended ablation studies will help determine the extra benefits of factors such as genotype and additional imaging modalities in clinical decision models.

Conclusion

This study aims to explore how to construct a model for predicting the overall survival rate of low-grade glioma (LGG) patients using pathological images as the primary data source, combined with basic clinical information. The study demonstrates that a model focusing on distinctive features in pathological images and incorporating patients’ age information can improve predictive accuracy and robustness, providing new insights and methods for clinical practice.