Abstract
Diffuse glioma is the most common group of malignant primary tumors of the central nervous system (CNS) in adults, and glioblastoma is the most aggressive form. Standard treatment comprises maximal safe surgical resection, followed by radiotherapy and chemotherapy. Still, the prognosis of this disease remains dismal with a median survival of only around 12 months in unselected, population based data. Extensive surgical resection is associated with longer survival, and the residual tumor (RT) volume and extent of resection (EOR) are important prognostic factors for these patients. In the presence of a RT, the EOR can be calculated as the ratio between pre-operative and surgically removed tumor tissue.
An exact calculation of the EOR relies on accurate measurements of the pre-operative tumor and RT volumes. The current standard method for estimating the RT volume in MR images is based on the product of the largest axial perpendicular diameters in 2D. This is a crude measure, and subject to high inter- and intra-rater variability. In the last decade, a substantial number of contributions have demonstrated the successful application of deep learning models for segmentation of glioblastoma from pre-operative MRI scans. In contrast, the analogous task of RT segmentation from early post-operative MRI (EPMR) scans has been less studied mainly due to the absence of publicly available datasets or benchmarks.
This thesis consists of four papers, revolving around automatic segmentation of RT from EPMR scans. In paper I (P1), we demonstrated the feasibility of RT segmentation from EPMR scans after surgical resection of glioblastoma, using current state-of-the-art deep learning models for pre-operative glioblastoma segmentation. The models were trained and validated on the largest dataset for early post-operative segmentation presented at the time, consisting of annotated multi-sequence EPMR scans from 956 patients, originating from 12 different hospitals in Europe and the U.S. The models' automatic RT segmentation and classification performances were shown to be on par with the human expert inter-rater variability.
In paper II (P2), the prognostic value of the automatic RT volume and EOR measurements was assessed and compared against the prognostic value of manual measurements, using Cox regression models. The results showed that both manually and automatically RT volumes were negative prognostic factors for overall survival. Similarly, both manual and automatic EOR models showed that patients with gross-total resection had longer overall survival than those with sub-total resection. Automatically measured EOR and RT volume had comparable prognostic properties to manually measured volumetric assessments.
In paper III (P3), targeted methods for improving the detection of small lesions were explored, focusing on sampling strategies, receptive field size and network depth. The presented experiments yielded only marginal improvements in voxel-wise segmentation performance and patient-wise sensitivity, but at the cost of a worse patient-wise specificity. Inter-rater variability of annotations was highlighted as a key challenge, and this became a motivation for further assessment of the impact of image and annotation quality.
In paper IV (P4), the impact of data quality on the models' segmentation performance was further studied. The image and annotation qualities of data originating from two hospitals were evaluated by one expert rater, and models were trained on different quality-based subsets of the dataset. Both image and annotation quality had a significant impact on model performance. Including images of various qualities promoted robustness and generalization of the models, without compromising on the performance on high quality images. High annotation quality was highlighted as a priority in dataset curation.