A Bayesian Deep Segmentation Framework for Glioblastoma Tumor Segmentation Using Follow-up MRIs

Kang-Lin Hsieh¹Tanjida Kabir¹Luis Nunez¹Yu-Chun Hsu¹Juan C Rodriguez Quintero¹Octavio Arevalo²Kangyi Zhao³Jay-jiguang Zhu¹Roy Francisco Riascos¹Xiaoqian Jiang¹Shayan Shams^1*

• ¹University of Texas Health Science Center at Houston, Houston, United States
• ²Louisiana State University Health Sciences Center, Shreveport, United States
• ³University of Pittsburgh, Pittsburgh, United States

Background: Glioblastoma (GBM) is the most common malignant brain tumor with an abysmal prognosis. Since complete tumor cell removal is impossible due to the infiltrative nature of GBM, accurate measurement is paramount for GBM assessment. Preoperative magnetic resonance images (MRIs) are crucial for initial diagnosis and surgical planning, while follow-up MRIs are vital for evaluating treatment response. The structural changes in the brain caused by surgical and therapeutic measures create significant differences between preoperative and follow-up MRIs. In clinical research, advanced deep learning models trained on preoperative MRIs are often applied to assess follow-up scans, but their effectiveness in this context remains underexplored. Our study evaluates the performance of these models on follow-up MRIs, revealing suboptimal results. To overcome this limitation, we developed a Bayesian deep segmentation model specifically designed for follow-up MRIs. This model is capable of accurately segmenting various GBM tumor sub-regions, including FLAIR hyperintensity regions, enhancing tumor areas, and non-enhancing central necrosis regions. By integrating uncertainty information, our model can identify and correct misclassifications, significantly improving segmentation accuracy. Therefore, the goal of this study is to provide an effective deep segmentation model for accurately segmenting GBM tumor sub-regions in follow-up MRIs, ultimately enhancing clinical decision-making and treatment evaluation. Methods: A novel deep segmentation model was developed utilizing 311 follow-up MRIs to segment tumor subregions. This model integrates Bayesian learning to assess the uncertainty of its predictions and employs transfer learning techniques to effectively recognize and interpret textures and spatial details of regions that are typically underrepresented in follow-up MRI data. Results: The proposed model significantly outperformed existing models, achieving DSC scores of 0.833, 0.901, and 0.931 for fluid attenuation inversion recovery hyperintensity, enhancing tumoral and non-enhancing central necrosis, respectively. Conclusions: Our proposed model incorporates brain structural changes following surgical and therapeutic interventions and leverages uncertainty metrics to refine estimates of tumor, demonstrating the potential for improved patient management.