Deep Learning of Imaging Phenotype and Genotype for Predicting Overall Survival Time of Glioblastoma Patients

Globally, the most common and deadly malignant brain tumor is glioblastoma (Glioblastoma, GBM). In recent years, research has continuously attempted to predict the overall survival time (OS) of GBM patients using machine learning techniques based on preoperative single-modality or multi-modality imaging phenotypes. Although these machine learning methods have made certain progress, most studies have not considered the inclusion of tumor genotype information in imaging-based OS prediction methods, which strongly indicates prognostic outcomes. To address this issue, Tang Zhenyu, Xu Yuyun, Jin Lei, and others published a research paper titled “Deep Learning of Imaging Phenotype and Genotype for Predicting Overall Survival Time of Glioblastoma Patients” in the June 2020 issue of “IEEE Transactions on Medical Imaging”. They proposed a new OS prediction method based on deep learning that derives features related to tumor genotype from preoperative multi-modality magnetic resonance imaging (MRI) data and inputs them into OS prediction.

The researchers proposed a multitask convolutional neural network (CNN) designed to accomplish both tumor genotype and OS prediction tasks. The network can improve OS prediction accuracy by learning features related to genotype prediction. In the experiments, a multimodal MRI brain dataset of 120 GBM patients was used, including up to four different genotypes/molecular biomarkers. The OS prediction accuracy of this method surpassed other state-of-the-art methods.

The background of this study aligns with clinical needs, targeting precise preoperative prognosis evaluation for personalized treatment of GBM patients. The institutions involved in this research include the Beijing Advanced Innovation Center for Big Data and Brain Computing of Beihang University, Zhejiang Provincial People’s Hospital, and Hangzhou Medical College, showcasing the strength of interdisciplinary collaboration.

The experimental results based on the single-center GBM biobank database, which includes multimodal MRI, genomic features, and OS information, demonstrate that this method is superior to existing radiomics-based and deep learning-based OS prediction methods.

Additionally, the study’s collected data underwent rigorous preprocessing, defining a uniform 3D multimodal MRI patch size to ensure coverage of the entire brain tumor for all patients. Detailed training and testing processes were used, along with data augmentation techniques to increase sample size and improve the model’s generalization ability.

The conclusions and significance of the research lie in the significant improvement in GBM prognosis accuracy by incorporating the tumor genotype prediction task, highlighting the advantages of deep learning in medical imaging analysis over traditional machine learning. By evaluating the impact of different genetic markers on OS prediction, the study found that the MGMT status and DWI are crucial genetic and imaging phenotype features for OS prediction. This finding not only reveals the close relationship between imaging phenotype, genotype, and clinical outcomes but also provides essential clinical guidance for GBM treatment planning.

The highlights of this study include the discovery of the relative importance of the MGMT gene marker and DWI, emphasizing the significance of genome-genome and/or genome-prognosis correlations. It underscores the close relationship between imaging phenotypes, genotypes, and clinical outcomes, providing valuable clinical guidelines for GBM treatment planning.

Furthermore, it is necessary to point out the study’s limitations, as only DWI and T1c modality data were used, and the exclusion of more modalities may limit further improvement in prediction accuracy. The incompleteness of the data might also have affected the model’s performance. Future research could incorporate intraoperative and postoperative images and tumor resection information to further update OS predictions.

This paper integrates advanced deep learning methods and medical imaging technology, providing a new high-accuracy prognostic prediction model for GBM patients, offering strong guidance for clinical practice and future research directions.