Keywords:
Artificial Intelligence, Neuroradiology brain, MR, Contrast agent-other, Cancer
Authors:
S. Bhardwaj1, C. Chong1, M. Agzarian1, S. Pati2, U. Baid2, S. Bakas2, M.-S. To1; 1Bedford Park, SA/AU, 2Pennsylvania, Philadelphia/US
DOI:
10.26044/ranzcr2021/R-0362
Methods and materials
Histologically-confirmed cases of glioblastoma managed under the neurosurgical department at Flinders Medical Centre (FMC, South Australia) between 2010 and 2019 were identified. We contributed 110 cases to the FeTS training. Pre-operative baseline magnetic resonance imaging (MRI) scans comprising volumes acquired as T1-weighted, T1-weighted after Gadolinium contrast enhancement (T1-ce), T2-weighted, and T2-weighted-Fluid-Attenuated Inversion Recovery (FLAIR) were extracted from FMC’s imaging server. Pre-processing of the imaging data consisted of image registration to a standard atlas [3] and skull-stripping [4]. Initial inference performed by the FeTS platform produced a tumor segmentation using a label fusion [5] of the current state of the art DL models [6]. These tumor segmentation labels do not reflect strict histological boundaries but are MRI sequence image-based boundaries.
Specifically, the DL model segmented the tumor into the following 3 tumor sub-regions and color labels (Figure 1):
- ET: Enhancing part of the tumor (yellow)
- ED: Edematous and infiltrated tissue around the tumor (green)
- NET: Necrotic core of the tumor (red)
These segmentations were then reviewed and manually refined using the ITK-SNAP application [6]. The following MRI sequences were used to estimate the tumor region boundaries:
- Using T1-ce sequence, hyperintense signal regions compared to T1 images were labelled as ET regions.
- To distinguish between CSF and edema, T2-FLAIR images were compared to T2 images, and hyperintense signal regions only on T2-FLAIR sequence were labelled as ED regions.
- Hypointense signal regions on the T1-ce within the hyperintense signal regions when compared to T1 and normal healthy white matter were labelled as NET regions.
Thus, using the image-based boundaries as depicted by the MRI scans, the various tumor sub-regions were checked and corrected for labelling on all the cases. The corrections were performed on axial sections of the MRIs. Segmentation labels were additionally reviewed by a Radiologist (with at least 6 years of post-fellowship experience). These corrected segmentation data were then fed into the FeTS tool.
Local model training was performed on an AMD 24-core Epyc server equipped with NVIDIA RTX 3090 graphics processing units (GPUs) and 24Gb of RAM. Model training was executed in both CPU-only, or GPU modes.