Impact of signal intensity normalization of MRI on the generalizability of radiomic-based prediction of molecular glioma subtypes

OBJECTIVES

Radiomic features have demonstrated encouraging results for non-invasive detection of molecular biomarkers, but the lack of guidelines for pre-processing MRI-data has led to poor generalizability. Here, we assessed the influence of different MRI-intensity normalization techniques on the performance of radiomics-based models for predicting molecular glioma subtypes.

METHODS

Preoperative MRI-data from n = 615 patients with newly diagnosed glioma and known isocitrate dehydrogenase (IDH) and 1p/19q status were pre-processed using four different methods: no normalization (naive), N4 bias field correction (N4), N4 followed by either WhiteStripe (N4/WS), or z-score normalization (N4/z-score). A total of 377 Image-Biomarker-Standardisation-Initiative-compliant radiomic features were extracted from each normalized data, and 9 different machine-learning algorithms were trained for multiclass prediction of molecular glioma subtypes (IDH-mutant 1p/19q codeleted vs. IDH-mutant 1p/19q non-codeleted vs. IDH wild type). External testing was performed in public glioma datasets from UCSF (n = 410) and TCGA (n = 160).

RESULTS

Support vector machine yielded the best performance with macro-average AUCs of 0.84 (naive), 0.84 (N4), 0.87 (N4/WS), and 0.87 (N4/z-score) in the internal test set. Both N4/WS and z-score outperformed the other approaches in the external UCSF and TCGA test sets with macro-average AUCs ranging from 0.85 to 0.87, replicating the performance of the internal test set, in contrast to macro-average AUCs ranging from 0.19 to 0.45 for naive and 0.26 to 0.52 for N4 alone.

CONCLUSION

Intensity normalization of MRI data is essential for the generalizability of radiomic-based machine-learning models. Specifically, both N4/WS and N4/z-score approaches allow to preserve the high model performance, yielding generalizable performance when applying the developed radiomic-based machine-learning model in an external heterogeneous, multi-institutional setting.

CLINICAL RELEVANCE STATEMENT

Intensity normalization such as N4/WS or N4/z-score can be used to develop reliable radiomics-based machine learning models from heterogeneous multicentre MRI datasets and provide non-invasive prediction of glioma subtypes.

KEY POINTS

• MRI-intensity normalization increases the stability of radiomics-based models and leads to better generalizability. • Intensity normalization did not appear relevant when the developed model was applied to homogeneous data from the same institution. • Radiomic-based machine learning algorithms are a promising approach for simultaneous classification of IDH and 1p/19q status of glioma.

Deep-learning-based reconstruction of undersampled MRI to reduce scan times: a multicentre, retrospective, cohort study

BACKGROUND

The extended acquisition times required for MRI limit its availability in resource-constrained settings. Consequently, accelerating MRI by undersampling k-space data, which is necessary to reconstruct an image, has been a long-standing but important challenge. We aimed to develop a deep convolutional neural network (dCNN) optimisation method for MRI reconstruction and to reduce scan times and evaluate its effect on image quality and accuracy of oncological imaging biomarkers.

METHODS

In this multicentre, retrospective, cohort study, MRI data from patients with glioblastoma treated at Heidelberg University Hospital (775 patients and 775 examinations) and from the phase 2 CORE trial (260 patients, 1083 examinations, and 58 institutions) and the phase 3 CENTRIC trial (505 patients, 3147 examinations, and 139 institutions) were used to develop, train, and test dCNN for reconstructing MRI from highly undersampled single-coil k-space data with various acceleration rates (R=2, 4, 6, 8, 10, and 15). Independent testing was performed with MRIs from the phase 2/3 EORTC-26101 trial (528 patients with glioblastoma, 1974 examinations, and 32 institutions). The similarity between undersampled dCNN-reconstructed and original MRIs was quantified with various image quality metrics, including structural similarity index measure (SSIM) and the accuracy of undersampled dCNN-reconstructed MRI on downstream radiological assessment of imaging biomarkers in oncology (automated artificial intelligence-based quantification of tumour burden and treatment response) was performed in the EORTC-26101 test dataset. The public NYU Langone Health fastMRI brain test dataset (558 patients and 558 examinations) was used to validate the generalisability and robustness of the dCNN for reconstructing MRIs from available multi-coil (parallel imaging) k-space data.

FINDINGS

In the EORTC-26101 test dataset, the median SSIM of undersampled dCNN-reconstructed MRI ranged from 0·88 to 0·99 across different acceleration rates, with 0·92 (95% CI 0·92-0·93) for 10-times acceleration (R=10). The 10-times undersampled dCNN-reconstructed MRI yielded excellent agreement with original MRI when assessing volumes of contrast-enhancing tumour (median DICE for spatial agreement of 0·89 [95% CI 0·88 to 0·89]; median volume difference of 0·01 cm3 [95% CI 0·00 to 0·03] equalling 0·21%; p=0·0036 for equivalence) or non-enhancing tumour or oedema (median DICE of 0·94 [95% CI 0·94 to 0·95]; median volume difference of -0·79 cm3 [95% CI -0·87 to -0·72] equalling -1·77%; p=0·023 for equivalence) in the EORTC-26101 test dataset. Automated volumetric tumour response assessment in the EORTC-26101 test dataset yielded an identical median time to progression of 4·27 months (95% CI 4·14 to 4·57) when using 10-times-undersampled dCNN-reconstructed or original MRI (log-rank p=0·80) and agreement in the time to progression in 374 (95·2%) of 393 patients with data. The dCNN generalised well to the fastMRI brain dataset, with significant improvements in the median SSIM when using multi-coil compared with single-coil k-space data (p<0·0001).

INTERPRETATION

Deep-learning-based reconstruction of undersampled MRI allows for a substantial reduction of scan times, with a 10-times acceleration demonstrating excellent image quality while preserving the accuracy of derived imaging biomarkers for the assessment of oncological treatment response. Our developments are available as open source software and hold considerable promise for increasing the accessibility to MRI, pending further prospective validation.

FUNDING

Deutsche Forschungsgemeinschaft (German Research Foundation) and an Else Kröner Clinician Scientist Endowed Professorship by the Else Kröner Fresenius Foundation.

Intrathecal IgM synthesis as a diagnostic marker in patients with suspected CNS lymphoma

For CNS lymphomas (CNSL), there is a high need for minimally invasive and easily obtainable diagnostic markers. Intrathecal IgM synthesis can easily be determined in routine CSF diagnostics. The aim of this study was to systematically investigate the diagnostic potential of intrathecal IgM synthesis in primary and secondary CNSL (PCNSL and SCNSL). In this retrospective study, patients with a biopsy-proven diagnosis of PCNSL or SCNSL were compared with patients with other neurological diseases in whom CNSL was initially the primary radiological differential diagnosis based on MRI. Sensitivity and specificity of intrathecal IgM synthesis were calculated using receiver operating characteristic curves. Seventy patients with CNSL were included (49 PCNSL and 21 SCNSL) and compared to 70 control patients. The sensitivity and specificity for the diagnosis of CNSL were 49% and 87%, respectively, for the entire patient population and 66% and 91% after selection for cases with tumor access to the CSF system and isolated intrathecal IgM synthesis. In cases with MRI-based radiological suspicion of CNSL, intrathecal IgM synthesis has good specificity but limited sensitivity. Because of its low-threshold availability, analysis of intrathecal IgM synthesis has the potential to lead to higher diagnostic accuracy, especially in resource-limited settings, and deserves further study.

Advancing noninvasive glioma classification with diffusion radiomics: Exploring the impact of signal intensity normalization

Background

This study investigates the influence of diffusion-weighted Magnetic Resonance Imaging (DWI-MRI) on radiomic-based prediction of glioma types according to molecular status and assesses the impact of DWI intensity normalization on model generalizability.

Methods

Radiomic features, compliant with image biomarker standardization initiative standards, were extracted from preoperative MRI of 549 patients with diffuse glioma, known IDH, and 1p19q-status. Anatomical sequences (T1, T1c, T2, FLAIR) underwent N4-Bias Field Correction (N4) and WhiteStripe normalization (N4/WS). Apparent diffusion coefficient (ADC) maps were normalized using N4 or N4/z-score. Nine machine-learning algorithms were trained for multiclass prediction of glioma types (IDH-mutant 1p/19q codeleted, IDH-mutant 1p/19q non-codeleted, IDH-wild type). Four approaches were compared: Anatomical, anatomical + ADC naive, anatomical + ADC N4, and anatomical + ADC N4/z-score. The University of California San Francisco (UCSF)-glioma dataset (n = 409) was used for external validation.

Results

Naïve-Bayes algorithms yielded overall the best performance on the internal test set. Adding ADC radiomics significantly improved AUC from 0.79 to 0.86 (P = .011) for the IDH-wild-type subgroup, but not for the other 2 glioma subgroups (P > .05). In the external UCSF dataset, the addition of ADC radiomics yielded a significantly higher AUC for the IDH-wild-type subgroup (P ≤ .001): 0.80 (N4/WS anatomical alone), 0.81 (anatomical + ADC naive), 0.81 (anatomical + ADC N4), and 0.88 (anatomical + ADC N4/z-score) as well as for the IDH-mutant 1p/19q non-codeleted subgroup (P < .012 each).

Conclusions

ADC radiomics can enhance the performance of conventional MRI-based radiomic models, particularly for IDH-wild-type glioma. The benefit of intensity normalization of ADC maps depends on the type and context of the used data.

Cluster-based prognostication in glioblastoma: Unveiling heterogeneity based on diffusion and perfusion similarities

BACKGROUND

While the association between diffusion and perfusion MRI and survival in glioblastoma is established, prognostic models for patients are lacking. This study employed clustering of functional imaging to identify distinct functional phenotypes in untreated glioblastomas, assessing their prognostic significance for overall survival.

METHODS

A total of 289 patients with glioblastoma who underwent preoperative multimodal MR imaging were included. Mean values of apparent diffusion coefficient (ADC) normalized relative cerebral blood volume (nrCBV), and relative cerebral blood flow (rCBF) were calculated for different tumor compartments and the entire tumor. Distinct imaging patterns were identified using Partition Around Medoids (PAM) clustering on the training dataset, and their ability to predict overall survival was assessed. Additionally, tree-based machine-learning models were trained to ascertain the significance of features pertaining to cluster membership.

RESULTS

Using the training dataset (231/289) we identified two stable imaging phenotypes through PAM clustering with significantly different overall survival (OS). Validation in an independent test set revealed a high-risk group with a median OS of 10.2 months and a low-risk group with a median OS of 26.6 months (p=0.012). Patients in the low-risk cluster had high diffusion and low perfusion values throughout, while the high-risk cluster displayed the reverse pattern. Including cluster membership in all multivariate Cox regression analyses improved performance (p≤ 0.004 each).

CONCLUSIONS

Our research demonstrates that data-driven clustering can identify clinically relevant, distinct imaging phenotypes, highlighting the potential role of diffusion and perfusion MRI in predicting survival rates of glioblastoma patients.

The 2021 WHO Classification for Gliomas and Implications on Imaging Diagnosis: Part 3-Summary of Imaging Findings on Glioneuronal and Neuronal Tumors

The fifth edition of the World Health Organization classification of central nervous system tumors published in 2021 reflects the current transitional state between traditional classification system based on histopathology and the state-of-the-art molecular diagnostics. This Part 3 Review focuses on the molecular diagnostics and imaging findings of glioneuronal and neuronal tumors. Histological and molecular features in glioneuronal and neuronal tumors often overlap with pediatric-type diffuse low-grade gliomas and circumscribed astrocytic gliomas (discussed in the Part 2 Review). Due to this overlap, in several tumor types of glioneuronal and neuronal tumors the diagnosis may be inconclusive with histopathology and genetic alterations, and imaging features may be helpful to distinguish difficult cases. Thus, it is crucial for radiologists to understand the underlying molecular diagnostics as well as imaging findings for application on clinical practice. EVIDENCE LEVEL: 3 TECHNICAL EFFICACY: Stage 3.

The 2021 WHO Classification for Gliomas and Implications on Imaging Diagnosis: Part 2-Summary of Imaging Findings on Pediatric-Type Diffuse High-Grade Gliomas, Pediatric-Type Diffuse Low-Grade Gliomas, and Circumscribed Astrocytic Gliomas

The fifth edition of the World Health Organization (WHO) classification of central nervous system tumors published in 2021 advances the role of molecular diagnostics in the classification of gliomas by emphasizing integrated diagnoses based on histopathology and molecular information and grouping tumors based on genetic alterations. This Part 2 review focuses on the molecular diagnostics and imaging findings of pediatric-type diffuse high-grade gliomas, pediatric-type diffuse low-grade gliomas, and circumscribed astrocytic gliomas. Each tumor type in pediatric-type diffuse high-grade glioma mostly harbors a distinct molecular marker. On the other hand, in pediatric-type diffuse low-grade gliomas and circumscribed astrocytic gliomas, molecular diagnostics may be extremely complicated at a glance in the 2021 WHO classification. It is crucial for radiologists to understand the molecular diagnostics and imaging findings and leverage the knowledge in clinical practice. EVIDENCE LEVEL: 3 TECHNICAL EFFICACY: Stage 3.

The 2021 WHO Classification for Gliomas and Implications on Imaging Diagnosis: Part 1-Key Points of the Fifth Edition and Summary of Imaging Findings on Adult-Type Diffuse Gliomas

The fifth edition of the World Health Organization (WHO) classification of central nervous system tumors published in 2021 advances the role of molecular diagnostics in the classification of gliomas by emphasizing integrated diagnoses based on histopathology and molecular information and grouping tumors based on genetic alterations. Importantly, molecular biomarkers that provide important prognostic information are now a parameter for establishing tumor grades in gliomas. Understanding the 2021 WHO classification is crucial for radiologists for daily imaging interpretation as well as communication with clinicians. Although imaging features are not included in the 2021 WHO classification, imaging can serve as a powerful tool to impact the clinical practice not only prior to tissue confirmation but beyond. This review represents the first of a three-installment review series on the 2021 WHO classification for gliomas, glioneuronal tumors, and neuronal tumors and implications on imaging diagnosis. This Part 1 Review focuses on the major changes to the classification of gliomas and imaging findings on adult-type diffuse gliomas. EVIDENCE LEVEL: 3. TECHNICAL EFFICACY: Stage 3.

Automated hippocampal segmentation algorithms evaluated in stroke patients

Deep learning segmentation algorithms can produce reproducible results in a matter of seconds. However, their application to more complex datasets is uncertain and may fail in the presence of severe structural abnormalities-such as those commonly seen in stroke patients. In this investigation, six recent, deep learning-based hippocampal segmentation algorithms were tested on 641 stroke patients of a multicentric, open-source dataset ATLAS 2.0. The comparisons of the volumes showed that the methods are not interchangeable with concordance correlation coefficients from 0.266 to 0.816. While the segmentation algorithms demonstrated an overall good performance (volumetric similarity [VS] 0.816 to 0.972, DICE score 0.786 to 0.921, and Hausdorff distance [HD] 2.69 to 6.34), no single out-performing algorithm was identified: FastSurfer performed best in VS, QuickNat in DICE and average HD, and Hippodeep in HD. Segmentation performance was significantly lower for ipsilesional segmentation, with a decrease in performance as a function of lesion size due to the pathology-based domain shift. Only QuickNat showed a more robust performance in volumetric similarity. Even though there are many pre-trained segmentation methods, it is important to be aware of the possible decrease in performance for the segmentation results on the lesion side due to the pathology-based domain shift. The segmentation algorithm should be selected based on the research question and the evaluation parameter needed. More research is needed to improve current hippocampal segmentation methods.

Assessing the added value of ADC, CBV and radiomic MR features for differentiation of pseudoprogression vs. true tumor progression in patients with glioblastoma

Background

Pseudoprogression (PsPD) is a major diagnostic challenge in the follow-up of patients with glioblastoma (GB) after chemo-radiotherapy (CRT). Conventional imaging signs and parameters derived from diffusion and perfusion-MRI have yet to prove their reliability in clinical practice for an accurate differential diagnosis. Here, we tested these parameters and combined them with radiomic features (RFs), clinical data, and MGMT-promoter methylation-status using machine- and deep-learning models to distinguish PsPD from PD.

Methods

In a single-center analysis, 105 patients with GB who developed a suspected imaging PsPD in the first 7 months after standard-CRT were identified retrospectively. Imaging data included standard MRI anatomical sequences, apparent diffusion coefficient (ADC) and normalized relative cerebral blood volume (nrCBV) maps. Median values (ADC, nrCBV) and RFs (all sequences) were calculated from deep-learning-based tumor segmentations. Generalized linear models with LASSO feature-selection and deep-learning models were built integrating clinical data, MGMT-methylation-status, median ADC and nrCBV values and RFs.

Results

A model based on clinical data and MGMT-methylation-status yielded an AUC=0.69 (95%CI 0.55-0.83) for detecting PsPD, and the addition of median ADC and nrCBV values resulted in a non-significant increase in performance (AUC=0.71,95%CI 0.57-0.85, p=0.416). Combining clinical/MGMT information with RFs derived from ADC, nrCBV and from all available sequences both resulted in significantly (both p&lt;0.005) lower model performances, with AUC=0.52 (0.38-0.66) and AUC=0.54 (0.40-0.68), respectively. Deep-learning imaging models resulted in AUCs≤0.56.

Conclusion

Currently available imaging biomarkers could not reliably differentiate PsPD from true tumor progression in patients with glioblastoma; larger collaborative efforts are needed to build more reliable models.