INTRODUCTION: Intracranial meningiomas are mostly benign intracranial lesions that arise from the dura mater and account for approximately 30% of primary intracranial tumors^1–3. Patients that are treated conservatively usually receive serial follow-up imaging in order to evaluate tumor growth⁴. Tumor progression pattern is known to be rather slow, multifocal, and multidirectional. Volumetric evaluation of meningiomas is therefore superior to traditional diametric methods when assessing tumor growth; however, it is proven to be time-consuming^5,6 and intra- and inter-reader variabilities in brain tumor segmentation are high, ranging up to 30%^7,8. Technical specifications and segmentation performances of the available deep-learning models differ^7,9.
We hypothesized training an established deep-learning-model architecture (3D-Deep-Convolutional-Neural-Network, DeepMedic) that already performed well in brain tumor segmentation^9–11 might allow for superior results and save resources than creating an entirely new deep learning model from scratch. The objective of our study was to train the above-mentioned architecture on manual meningioma segmentations and then validate its performance on an independent group of meningiomas from routine multiparametric MR imaging data.

METHODS: Our retrospective study was approved by the local institutional review board. Only patients with a complete MR-dataset, consisting of T1-/T2-weighted, T1-weighted contrast-enhanced [T1CE], and FLAIR as well as histopathological specimen were included. The final study population consisted of 125 patients classified according to the current World Health Organization (WHO) guideline¹². Patients were split into two groups: (i) training data for the deep learning model (training-group, n=70) and (ii) comparison between manual and automated segmentations (validation group, n=55). MR images were acquired on different scanner types (from all major vendors) from referring institutions (n=86) and our own institution (n=39), ranging from 1.0 to 3.0 Tesla.
Manual segmentations were performed using IntelliSpace Discovery (Philips Healthcare). In T1CE, contrast-enhancing tumor was segmented. In FLAIR, solid tumor and surrounding edema were segmented. For evaluation, we applied a previously implemented method¹¹ utilizing the following defined tumor volumes: (i) contrast-enhancing-tumor in T1CE and (ii) total-lesion-volume as the union of target volumes in T1CE and FLAIR, including solid contrast-enhancing tumor parts and surrounding edema. Manual segmentations in the training group were performed by two radiologists separately, resulting in two datasets of segmentations for each patient. To define optimal ground truth tumor volumes for the validation group, two additional radiologists segmented tumor volumes in consensus.
DeepMedic-architecture⁹ was used, including 3D-convolutional-neural-network for segmentation and a 3D-post-processing-step to remove false-positives. For training of the deep-learning-model, the two sets of manual segmentations from each reader in T1CE and FLAIR were used (totaling 140 image sets each including T1CE and FLAIR segmentations). For automated segmentation imaging data was pre-processed including registration, skull stripping, resampling, and normalization.
The dice-coefficient was used to assess similarity between manual ground truth segmentations (S1) and automated deep-learning based segmentations (S2) and in the validation group, DSC(S1,S2)=(2|S1∩S2|)÷(|S1|+|S2|). The dice-coefficient was also used to evaluate interreader-variabilities between the manual segmentations of the two readers of the training group¹³. Wilcoxon signed rank test was used to evaluate any statistical difference.

RESULTS: The 125 meningioma patients consisted of 97 WHO grade I meningioma and 28 patients with WHO grade II meningioma.
Training group: Manual segmentations for contrast-enhancing-tumor averaged 31.5±29.3 cm³ for the first and 31.1±28.9 cm³ for the second reader as well as 57.4±56.4 cm³ and 59.6±59.2 cm³ for total-lesion-volume. The dice-coefficients were 0.92±0.07 for contrast-enhancing-tumor and 0.88±0.05 for total-lesion-volume.
Validation group: Manual segmentations were 30.7±25.1 cm³ for contrast-enhancing-tumor volume and 71.3±66.0 cm³ for total-lesion-volume. Automated segmentations were comparable without significant differences, being 30.7±24.1 cm³ (p=0.95) in contrast-enhancing-tumor and 71.4±65.0 cm³ (p=0.94) for total-lesion-volume. The average dice-coefficients for comparing automated and manual segmentations were 0.82±0.12 (range: 0.54-0.97) for contrast-enhancing-tumor and 0.91±0.08 (range: 0.21-0.93) for total-lesion-volume. Automated segmentation performed significantly better in contrast-enhancing-tumor volume than the total-lesion-volume (0.91±0.08 vs. 0.82±0.12, p<0.001). There was no significant difference in dice-coefficients between grade I and II meningiomas or between skull base and convexity meningiomas.

DISCUSSION: Our trained deep-learning-model segmented contrast-enhancing-tumor and total-lesion-volume very accurately. The accuracy was comparable to the interreader-variabilities of the two readers from the training group. Further, our results are either comparable or superior to other recently published studies addressing automated brain tumor segmentation^{7,9–11,14–16} and general segmentation accuracies accounting for intra- and interreader-variabilities^7,8.
Reliable automated meningioma segmentation as demonstrated in our study should relevantly increase availability of volumetric data as time consuming manual segmentations could be omitted. Further, automated segmentation should improve reproducibility of measurements⁷. Accessibility of volumetric data is clinically warranted for several reasons. Monitoring of conservatively treated meningioma could be improved as tumor growth can be detected with higher sensitivity by a volumetric assessment compared to conventional diameter methods^5,6. Image-based tumor characterization by quantitative image analysis, e.g. radiomics-based or (deep) machine learning approaches, heavily rely on volumetric image segmentations^7,17–19. Further, extent of peritumoral edema has a decisive impact on the clinical outcome as well as intraoperative performance. Assessment of edema by the total-lesion-volume therefore appears warranted²⁰.

CONCLUSION: Automated meningioma segmentation by our trained deep learning is highly accurate and produces reliable results, which are comparable to the interreader-variabilities from manual readers.

Acknowledgements

No acknowledgement found.