PaperlayerView exact revision on arXiv

Source-linked AI summary

The ASNR-MICCAI Brain Tumor Segmentation (BraTS) Challenge 2023: Intracranial Meningioma

Dominic LaBella, Maruf Adewole, Michelle Alonso-Basanta, Talissa Altes, Syed Muhammad Anwar, Ujjwal Baid, Timothy Bergquist, Radhika Bhalerao, Sully Chen, Verena Chung, Gian-Marco Conte, Farouk Dako, James Eddy, Ivan Ezhov, Devon Godfrey, Fathi Hilal, Ariana Familiar, Keyvan Farahani, Juan Eugenio Iglesias, Zhifan Jiang, Elaine Johanson, Anahita Fathi Kazerooni, Collin Kent, John Kirkpatrick, Florian Kofler, Koen Van Leemput, Hongwei Bran Li, Xinyang Liu, Aria Mahtabfar, Shan McBurney-Lin, Ryan McLean, Zeke Meier, Ahmed W Moawad, John Mongan, Pierre Nedelec, Maxence Pajot, Marie Piraud, Arif Rashid, Zachary Reitman, Russell Takeshi Shinohara, Yury Velichko, Chunhao Wang, Pranav Warman, Walter Wiggins, Mariam Aboian, Jake Albrecht, Udunna Anazodo, Spyridon Bakas, Adam Flanders, Anastasia Janas, Goldey Khanna, Marius George Linguraru, Bjoern Menze, Ayman Nada, Andreas M Rauschecker, Jeff Rudie, Nourel Hoda Tahon, Javier Villanueva-Meyer, Benedikt Wiestler, Evan Calabrese

arXiv:2305.07642v1cs.CVcs.AIcs.LGstat.ML

TL;DR

Automated, objective quantitative tools for non-invasive meningioma assessment on mpMRI are lacking, despite MRI’s role in diagnosis and treatment planning. The BraTS 2023 challenge defines three segmentation targets and a benchmark using an expert-annotated mpMRI dataset. The authors present the challenge as a basis for future models intended to support monitoring, planning, and non-invasive assessment of meningiomas.

  • Problem

    Automated, objective quantitative tools for non-invasive meningioma mpMRI assessment are lacking, although MRI is used for diagnosis, treatment planning, and longitudinal monitoring.

    Source: §1 · p. 2

  • Method

    The BraTS 2023 challenge organizes automated segmentation of enhancing tumor, non-enhancing tumor core, and surrounding non-enhancing T2/FLAIR hyperintensity using an expert-annotated meningioma mpMRI dataset.

    Source: §1 · p. 2, §2.2 · p. 3, §3.1 · p. 6

  • Results

    The challenge provides a community standard and benchmark for automated intracranial meningioma segmentation based on an expert-annotated multilabel mpMRI dataset.

    Source: §4 · p. 9

  • Takeaways & Limitations

    The authors expect challenge-developed models to aid objective tumor monitoring, surgical and radiotherapy planning, and non-invasive assessment of clinically relevant tumor characteristics.

    Source: §4 · p. 9

    Preprocessing loses radiographic information and changes image data through resampling, while expert judgments of meningioma boundaries may differ between observers.

    Source: §3.3 · pp. 8–9

Abstract

from arXiv · show

Meningiomas are the most common primary intracranial tumor in adults and can be associated with significant morbidity and mortality. Radiologists, neurosurgeons, neuro-oncologists, and radiation oncologists rely on multiparametric MRI (mpMRI) for diagnosis, treatment planning, and longitudinal treatment monitoring; yet automated, objective, and quantitative tools for non-invasive assessment of meningiomas on mpMRI are lacking. The BraTS meningioma 2023 challenge will provide a community standard and benchmark for state-of-the-art automated intracranial meningioma segmentation models based on the largest expert annotated multilabel meningioma mpMRI dataset to date. Challenge competitors will develop automated segmentation models to predict three distinct meningioma sub-regions on MRI including enhancing tumor, non-enhancing tumor core, and surrounding nonenhancing T2/FLAIR hyperintensity. Models will be evaluated on separate validation and held-out test datasets using standardized metrics utilized across the BraTS 2023 series of challenges including the Dice similarity coefficient and Hausdorff distance. The models developed during the course of this challenge will aid in incorporation of automated meningioma MRI segmentation into clinical practice, which will ultimately improve care of patients with meningioma.

1 Introduction

Meningioma care needs better non-invasive MRI assessment: conventional features may not reflect grade or clinical course, and automated segmentation research has focused mainly on gliomas. The BraTS 2023 challenge sets out to develop multi-compartment MRI segmentation and describe its data preparation and evaluation.

1 Introduction

  • No reliable non-invasive method currently identifies meningioma grade, assesses aggressiveness, or predicts recurrence and survival; volume and peritumoral edema may not represent grade or clinical course.The paper identifies improved radiographic assessment as a need for guiding patient-specific treatment strategy.Source: §1 · p. 2
  • Meningioma segmentation research faces extra-axial location, multiplicity, and skull-base involvement, while most tumor segmentation studies have focused on gliomas.Because meningiomas are commonly diagnosed by imaging alone, MRI has added importance for treatment planning; preoperative aggressiveness prediction could guide surgical strategy.Source: §1 · p. 2
  • The BraTS 2023 challenge aims to develop automated multi-compartment meningioma MRI segmentation and describes its annotation protocol, tasks, and evaluation metrics.The manuscript also outlines challenge objectives, constraints, and future directions under consideration.Source: §1 · p. 2

2 Materials & Methods

The paper defines intracranial meningioma by location within the cranial vault and summarizes its WHO-grade distribution and diagnostic features. Imaging findings vary across subtypes, and reliable imaging features for distinguishing grade or aggressiveness are lacking.

2.1 Defining intracranial meningioma

  • About 80% of meningiomas are WHO grade 1, approximately 18% grade 2, and only about 2% grade 3.The challenge definition includes any grade or subtype occurring within the cranial vault.Source: §2.1 · pp. 2–3
  • WHO grade 1 subtypes include meningothelial, fibroblastic, transitional, psammomatous, angiomatous, microcystic, secretory, metaplastic, and lymphoplasmacyte rich meningiomas.Typical histology includes psammoma bodies, cellular whirls, and calcifications.Source: §2.1 · p. 2
  • Grade 2 criteria include ≥4 mitoses per high-power field, brain invasion, or at least three listed histologic features; grade 3 includes ≥10 mitoses/HPF and/or specified malignant features.The grade 2 features include hypercellularity, small cells, prominent nucleoli, patternless or sheet-like growth, and spontaneous necrosis; grade 3 features include loss of usual growth pattern or multifocal necrotic foci.Source: §2.1 · pp. 2–3
  • Radiographic findings vary among meningioma subtypes, and no reliable imaging features currently distinguish meningioma grade or aggressiveness.Source: §2.1 · p. 3

2.2 Defining Meningioma MRI Sub-Components

The challenge assigns three nonoverlapping MRI labels to meningioma-related findings. Their boundaries distinguish enhancing tumor, non-enhancing tumor core, and surrounding T2/FLAIR hyperintensity from unrelated signal abnormality.

2.2 Defining Meningioma MRI Sub-Components

  • The challenge draws on prior compartment-based BraTS work and ANOCEF consensus, which defines enhancing gross tumor to include contrast-enhancing lesions, thickened meninges, and directly invaded bone.The paper notes that prospective-study guidelines for intracranial meningioma target-volume delineation had not been published.Source: §2.2 · p. 3
  • MRI appearances include T2-low-signal non-enhancing calcification, FLAIR-hyperintense peritumoral edema, en-plaque enhancing dural sheets, and dural tails extending away from lesions.The paper reports peritumoral edema in 60% of cases and dural tails in approximately 72%; dural tails are absent in intracranial Schwannoma mimics.Source: §2.2 · p. 3
  • The three labels are enhancing tumor, non-enhancing tumor core, and surrounding non-enhancing T2/FLAIR hyperintensity (SNFH), which is entirely brain parenchyma and not expected to contain tumor cells.Enhancing tumor plus non-enhancing tumor core form the tumor core; SNFH represents irritated, inflamed, and/or edematous tissue adjacent to tumor.Source: §2.2 · p. 3
  • Enhancing tumor includes contrast-enhancing meningioma, focally thickened meninges including dural tail, and en-plaque meningiomas, approximating active, viable tumor.Source: §2.2 · p. 3
  • Non-enhancing tumor core includes calcification, hyperostosis, necrosis, degeneration, and other atypical non-enhancing tumor findings; together with enhancing tumor, it approximates tissue typically removed in gross total resection.Source: §2.2 · p. 3
  • SNFH excludes non-tumor-related T2/FLAIR signal abnormality, commonly associated with chronic microvascular ischemic white matter changes or other vascular pathology.Source: §2.2 · p. 4

2.3 Data Description

The dataset contains pretreatment, preoperative mpMRI studies of meningiomas, excluding cases with other intracranial tumors. Each included case has four specified MRI series.

2.3 Data Description

  • Included studies were preoperative and pretreatment, with one or more meningiomas in the field of view; studies with another intracranial tumor were excluded.Included mpMRI series are pre-contrast T1-weighted, post-contrast T1-weighted, T2-weighted, and T2-weighted FLAIR.Source: §2.3 · p. 4

2.4 Participating sites

Participating centers contributed preoperative mpMRI cases accepted by pathology or clinical/radiographic diagnosis, with case-collection practices and imaging protocols varying across and within sites.

2.4 Participating sites

  • Cases were identified through histopathology after resection or biopsy, or through a formal clinical and radiographic diagnosis of meningioma.Imaging-based diagnosis was included because meningiomas may be observed or treated without definitive tissue diagnosis.Source: §2.4 · p. 4
  • Case inclusion and collection methods were chosen independently by participating sites, and imaging parameters varied considerably within and between sites.Contributors were not required to disclose collection methods or MRI protocol information.Source: §2.4 · p. 4
  • The reported site contributions include only preoperative meningiomas within the intracranial brain mask and judged radiographically or pathologically consistent with meningioma.Source: Table 1 · p. 5

2.5 Image data preprocessing

The mpMRI preprocessing pipeline converted, aligned, resampled, and skull-stripped images, excluding any extracranial tumor portions in the process.

2.5 Image data preprocessing

  • Preprocessing converted DICOM series to NIfTI and co-registered T1-weighted, T2-weighted, T2-FLAIR, and T1Gd images to SRI24 atlas space with uniform 1 mm^3 resampling.Source: §2.5 · p. 5
  • Automated deep convolutional neural network skull-stripping was applied, implicitly excluding any extracranial tumor portions.Source: §2.5 · p. 5
  • Skull-stripping was retained to preserve patient anonymity by preventing face reconstruction and to maintain consistency with other BraTS challenges.Source: §2.5 · p. 5

2.6 Pre-segmentation

A periodically retrained automated model pre-segmented cases before volunteer correction, followed by neuroradiologist review and iterative refinement when needed.

2.6 Pre-segmentation

  • The initial nnU-Net pre-segmentation model was trained on 73 manually labeled studies from one institution, all involving surgically resected meningiomas.The authors note this could bias performance toward poorer results on non-surgical meningiomas.Source: §2.6 · p. 6
  • The model was periodically retrained on manually corrected cases from additional sites, including non-surgical cases, to improve generalizability and reduce pre-segmentation bias.Source: §2.6 · p. 6

2.7 Manual corrections

  • Volunteer annotators received segmentation instructions and common-error examples, and fellowship-trained neuroradiologists reviewed every corrected case, returning inadequate labels for refinement.Annotators ranged from medical students to neuroradiologists with more than 10 years of experience.Source: §2.7 · p. 6

2.8 Common errors of automated segmentation

Annotators were given examples of recurring automated-label errors identified through subjective review of pre-segmented cases to help improve inter-observer variability.

2.8 Common errors of automated segmentation

  • The error examples were identified through subjective review of pre-segmented cases and supplied to annotators to improve inter-observer variability.Source: §2.8 · p. 6
  • Small meningiomas could receive a spurious thin rim of surrounding SNFH, while small convexity tumors could be incompletely or entirely missed, especially amid multiple tumors.Source: §2.8 · p. 6
  • Non-enhancing tumor core could be incompletely segmented or misassigned, including exophytic hyperostosis, cystic spaces, and intrinsic T1-hyperintense areas.Source: §2.8 · p. 6
  • Chronic microvascular ischemic white-matter changes could be incorrectly included in the SNFH label as tumor-related T2/FLAIR abnormality.Source: §2.8 · p. 6

3 Discussion

The challenge’s diverse, multi-site training data may improve model generalizability, with potential clinical uses in treatment planning, longitudinal assessment, and future prediction of tumor characteristics. Preprocessing and inter-observer variability constrain the challenge, while future iterations could expand its scope to additional tasks and patient settings.

3.1 Potential benefits of the challenge

  • Multi-site, heterogeneous mpMRI data mirror underlying data variation and may improve segmentation models’ generalizability at new sites.The passage identifies acquisition and processing protocols across sites as sources of batch effects.Source: §3.1 · pp. 6–7

3.2 Clinical relevance

  • Automated segmentations could support radiation planning, serial tumor-volume assessment, operative planning, and future prediction of meningioma characteristics.Potential future targets include grade, subtype, aggressiveness, and therapy response.Source: §3.2 · pp. 7–8

3.3 Limitations of the challenge

  • Preprocessing can remove extracranial tumor extension and alter native image resolution, while inter-observer variability in labels may constrain model performance.The authors note that multi-observer consensus segmentation is labor- and time-intensive.Source: §3.3 · pp. 8–9

3.4 Goals for future challenges

  • Future challenges could add grade prediction, tumor detection and measurement, post-treatment and extracranial cases, and recurrence-versus-treatment-change classification.The latter expansions would require new inclusion criteria and additional data.Source: §3.4 · p. 9

4 Conclusion

The challenge will establish a community benchmark for automated intracranial meningioma segmentation using a large expert-annotated multilabel mpMRI dataset, with potential applications in monitoring, planning, and non-invasive tumor characterization.

4 Conclusion

  • The challenge will provide a community standard and benchmark for automated intracranial meningioma segmentation using the largest expert-annotated multilabel mpMRI dataset to date.The models are intended to aid objective tumor monitoring, surgical and radiotherapy planning, and non-invasive assessment of clinically relevant tumor characteristics.Source: §4 · p. 9

Funding

The publication received partial NIH support, and its content does not represent the NIH’s official views.

Funding

  • The publication was partly supported by NIH awards NCI K08CA256045 and NCI/ITCR U01CA242871; its content does not represent NIH’s official views.Source: Funding · p. 9

Related papers

Loading 2305.07642v1…