William (Wiley) Winters

Determining the effectiveness of brain tumor treatments from post-treatment MRI images is an area where little research has been conducted. Current studies primarily focus on tumor detection, where MRI images clearly distinguish between healthy and tumorous tissue. The MR Images used in these studies clearly delineate between healthy and tumor images. For these cases, a Convolutional neural network (CNN) can be used to effectively classify these images as tumor-positive or negative. However, post-treatment imaging presents a challenge due to the obscuration of tumor cavities by edema and disrupted blood-brain barriers (BBB). As a result, CNNs trained on well-defined tumor images often perform poorly on post-treatment images, leading to incorrect classifications. This study aims to address this limitation by training a SegResNet from MONAI on a BraTS-style dataset for image segmentation. The trained model will be applied to a post-treatment dataset of glioma patients to segment the images and determine the presence or absence of tumor tissue. Accurate segmentation will enable the assessment of residual tumor tissue, which is crucial for developing informed treatment plans. By focusing on this specific area, this research seeks to improve the accuracy of post-treatment brain tumor diagnosis and follow-up care.

Brain tumors substantially impair patients’ quality of life owing to effects on cognition, personality, occupational function, and, if left untreated, survival. Magnetic resonance imaging (MRI) is the cornerstone of diagnosis, enabling detection, characterization, and longitudinal monitoring of tumors. Despite advances in medical image analysis, automated brain tumor segmentation remains challenging due to inter- and intra-tumor heterogeneity in location, morphology, size, and signal intensity. Tumors often exhibit indistinct margins, heterogeneous tissue composition, and complex cystic or necrotic components [1]. Although manual segmentation by certified neuroradiologists is feasible, it is labor-intensive, time-consuming, and subject to inter-rater variability [2]. These difficulties are notably amplified in post-operative MRI, where surgical cavities, peritumoral edema, and disrupted blood–brain barriers (BBB) obscure residual or recurrent tumor tissue, complicating accurate delineation [3]. To address this clinical challenge, this study proposes an enhanced U-Net architecture trained on University of Missouri Post-operative Glioma dataset to improve segmentation of post-treatment brain tumors. The trained model will be applied to a random sample of post-operative MRI scans, and its predictions will be evaluated against deep-learning annotations that have been reviewed by board certified neuroradiologists. Quantitative performance will be assessed using the Dice Similarity Coefficient (DSC) and the Jaccard similarity coefficient (IoU), standard metrics used for measuring spatial overlap between predicted and ground-truth segmentations.