Session Item

For glioblastoma (GBM), an isotropic margin of 20 mm from the contrast enhancing tumor is used because (i) conventional MRI and interpretation methods have limitations as to specify the location of non-enhancing infiltrative GBM, and (ii) recurrence commonly occur within 20 mm from the primary tumor. However, inter- and intra-observer variability and the diffuse radiological signature of GBM adds to the challenge of optimal visual interpretation and accurate delineation of GBM. A precise method is needed to specify the radiotherapy target and avoid unnecessary patient neurocognitive side effects.

The objective of this study was to investigate if a deep learning model for segmenting GBM on multispectral MRI correlated to manual delineations, and thereby potentially be feasible as an oncologist support tool. Secondly, we aim to investigate if the 20mm margin is justifiable in respect of geometrical appearance of recurrent GBM harboring unmethylated O⁶-methylguanine-DNA methyltransferase (MGMT) promotor.

Longitudinal image sets of multispectral post-operative MRI from six patients with (i) primary- and (ii) recurrent GBM were analyzed. All patients harbored unmethylated MGMT promoter, indicating high radioresistance and poor prognosis. We used a deep learning pre-trained U-Net based algorithm called HD-GLIO, trained on 3220 GBM image sets labeled by neuro-radiologists. The enhancing core (automatic GTV) and non-enhancing GBM compartments (automatic CTV) were derived and compared to the clinical GTVs and CTVs, respectively. The image series of the primary tumor was rigidly co-registered with the image series of the recurrent GBM for volumetric comparison between the clinical CTV and the recurrent tumor.

The automatic GTV overlapped with the clinical (true positive) by overall median 90% (range=33-100%) (p < 0.005), with a mean size=43% compared to the clinical GTVs. Median Dice coefficient=65% (range=15-77%) (p < 0.005).

The automatic CTV-segments overlapped with the clinical (true positive result) by median=98% (range=93-100%), mean size=56% compared to the clinical CTV. Median Dice coefficient=59% (range 15-77%) (p < 0.05).

The recurrent GBM overlapped with the primary clinical CTV by median 0% true positive (range=0-79%), and median 0% Dice coefficient (range 0-13%) (p < 0, 5).

We found a high degree of overlap in detecting the spatial area of malignant tissue using HD-GLIO, demonstrating the method feasible to employ as a support tool for oncologists.  Based on these findings, we have initialized a more extensive study.

We found a low degree of overlap between CTV-margin and recurrent tumor. This could suggest radiation treatment response, but considering the known tumor radioresistance it could also mean that irradiating this large margin of potentially normal tissue cells was only contributing to side effects. We suggest more research is performed to investigate recurrence patterns in relation to irradiated tissue volumes on this patient group.