UCSF

Current encoder-decoder convolutional neural networks (CNN) used in automated glioma lesion segmentation and volumetric measurements perform well on newly diagnosed lesions that have not received any treatment. However, there are challenges in generalizability for patients after treatment, including at the time of suspected recurrence. This results in decreased translation to clinical use in the post-treatment setting where it is needed the most. A potential reason is that these deep learning models are primarily trained on a singular curated dataset and demonstrate decreased performance when they are tested in situations with unseen variations to disease states, scanning protocols or equipment, and operators. While using a highly curated dataset does have the benefit of standardizing comparison of models, it comes with some significant drawbacks to generalizability. The primary source of images used to train current models for glioma segmentation is the BraTS (Multimodal Brain Tumor Image Segmentation Benchmark) dataset. The image domain of the BraTS dataset is large, including high- and low-grade tumors, varying acquisition resolution, and scans from multi-center studies. Despite this, it may still lack sufficient feature representation in the target clinical imaging domain. Here we address generalizability to the disease state of post-treatment glioma. The current BraTS dataset consists entirely of images obtained from newly diagnosed patients who have not undergone surgical resection, received adjuvant treatment, or shown significant disease progression, all of which can greatly alter the characteristics of these lesions. To improve the clinical utility of deep learning models for glioma segmentation, they must accommodate variations in signal intensity that may arise as a result of resection, tissue damage (treatment induced or otherwise), or progression. We compared models trained on either BraTS data, UCSF acquired post-treatment glioma data, UCSF acquired newly diagnosed glioma data, and various combinations of these data, to determine the effect of including images with features unique to treated gliomas into training the networks on segmentation performance in the post-treatment domain. Although an absolute threshold training inclusion value for generalization of segmentation networks to post-treatment glioma patients has not been established, we found that with 200 total training volumes, models trained with greater than or equal to 30% of the training images from patients with prior treatment received the greatest performance gains when testing in this domain. Additionally, we found that after this threshold is met, additional images from newly diagnosed patients did not negatively impact segmentation performance on patients with treated gliomas. We also developed a pre-processing pipeline and implemented a loss penalty term that incorporates cavity distance relationships to the tumor into weighting a cross entropy loss term. The aim of this was to bias the network weights to morphological features of the image relevant to pathologies that are prevalent post-treatment. This may either be used as an initialization for training with an available larger dataset such as BraTS or used to finetune a transferred network that has not seen sufficient post-treatment glioma images during training in order to allow domain adaptation with fewer training data from this disease state. Preliminary results show qualitatively more desirable segmentations of tumor lesions with respect to cavities and small disconnected components in selected examples that are worthy of further analysis with alternate training configurations, more focused performance assessments, and larger cohorts. Here, we will evaluate these techniques as potential solutions to improve the generalizability of CNN tumor segmentation to post- treatment glioma, as well as provide a framework for further data augmentation based on augmenting the boundary of these lesions.