RPS 1011b - Radiomics and deep learning in neuroimaging

Purpose:

Volumetric assessment of meningiomas represents a valuable tool for treatment planning and the evaluation of tumour growth as it allows for a more precise assessment of tumour size than conventional diameter methods. Segmentation data can be used for consecutive quantitative or radiomics analysis allowing for improved tumour characterisation. However, manual segmentations are time-consuming and therefore usually not performed. In this study, we established a dedicated meningioma deep learning model based on routine MRI-data and evaluated its performance for automated segmentation.

Methods and materials:

MRI datasets (T1-/T2-weighted, T1-weighted contrast-enhanced [T1CE], and FLAIR) of 126 patients with intracranial meningiomas (grade I: 97, grade II: 29) were included. Preprocessing of imaging data included registration, skull-stripping, resampling, and normalisation. Target volumes for manual and automated segmentations included contrast-enhancing tumour volume in T1CE and the total lesion volume (union of lesion volume in T1CE and FLAIR [including solid tumour parts and surrounding oedema]). For automated segmentation, an established deep learning model architecture (3D-Deep-Convolutional-Neural-Network, DeepMedic, BioMedIA) operating on all four MR-sequences was trained using manual segmentations of two independent readers from 70 patients (training-group). The trained deep learning model was then validated on the 56 remaining patients (validation-group) by comparing automated to ground-truth manual segmentations, which were performed by two additional readers in consensus.

Results:

In the validation-group, comparison of the deep learning model and manual segmentations revealed average Dice coefficients of 0.91±0.08 for the contrast-enhancing tumour volume and 0.82±0.12 for the total lesion volume. In the training-group, inter-reader variabilities of the two manual readers were 0.92±0.07 for the contrast-enhancing tumour volume and 0.88±0.05 for the total lesion volume.

Conclusion:

Deep learning-based automated segmentation yielded high segmentation accuracy, comparable to manual inter-reader variabilities.

Limitations:

Due to the retrospective study design, no evaluation of the actual clinical benefit could be applied.

Ethics committee approval

The study was approved by the institutional review board.

Funding:

No funding was received for this work.

Purpose:

To extract radiomics first-order distribution and texture features from post-surgical enhancing tumours and explore their prognostic value in patients with brain glioblastomas.

Methods and materials:

We retrospectively analysed 160 consecutive patients with glioblastoma multiforme treated with surgery and radiotherapy plus concomitant and adjuvant temozolomide from 2009-2017. Censored patients and those with images acquired later than 7 days from surgery were excluded. Subtraction of the T1-weighted and contrast-enhanced T1-weighted MRI was performed to obtain the enhancement map; the area of residual enhancement was semi-automatically segmented to obtain the enhancing mask. 96 radiomics features were extracted from the contrast-enhanced T1-weighted images (texture and first-order statistics). The population was split into training (80%) and test (20%) subpopulations. Feature selection was performed by minimum-redundancy-maximum-relevance and stepwise regression on the training set. Logistic regression and ROC curve analyses were performed to evaluate the prognosis into high/low survival (threshold at median survival of 16 months).

Results:

124 patients were included in the final analysis. A total of 10 radiomics features, including minimum intensity, contrast grey-level co-occurrence matrix, and inverse difference grey-level co-occurrence matrix features were selected. These features predicted short and long survival with an AUC=0.72 (95% CI 0.62-0.82, p<0.001) in the training and 0.71 (95% CI 0.49-0.92, p=0.044) in the test subpopulation.

Conclusion:

The selected radiomics features quantifying tumour intensity and heterogeneity are able to predict longer or shorter patient survival in both the training and test subpopulations with similar/fair accuracy.

Limitations:

In this retrospective study, the variability of image acquisition parameters was unavoidable. Manual segmentations might present interobserver variability. An external validation may confirm the applicability of the proposed method.

Ethics committee approval

Informed consent was waived by the Bellvitge University Hospital research ethics committee.

Funding:

RPL is supported by La-Caixa-Foundation and Prostate-Cancer-Foundation.

Purpose:

To study whether MR-based radiomic features can be used to assess the response to antituberculous treatment.

Methods and materials:

We included 24 patients with intracranial tuberculomas diagnosed by either culture or histopathology who had a pre- and post-treatment brain MRI with contrast performed at our institute from July 2009-July 2019. Patients with coexisting demyelination, tumour, or history of surgery were excluded. 16 patients showed a treatment response while 8 showed no response. A textural analysis was performed using Lifex and 44 textural parameters were extracted. The region of interest (ROI) was manually drawn on post-contrast FLAIR images. An independent Sample's t-test was employed to study statistically different parameters in both groups. Logistic regression was performed to develop a model for predicting the response to treatment on the basis of MR radiomic features in differentiating patients with intracranial tuberculoma.

Results:

MR radiomic parameters, histogram skewness and GLCM correlation, showed a statistically significant difference in patients who showed improvement versus those who did not (χ2=11.517, p=0.003). The model explained 52.9% (Nagelkerke R2) of variance in predicting the response to treatment and correctly classified 83.3% of cases. ROC curve analysis for histogram skewness showed an area under the curve of 0.766 (p=0.037 and 95% CI =0.577-0.954).

Conclusion:

MR textural parameters including histogram skewness and GLCM correlation can be used as imaging biomarkers to differentiate patients with intracranial tuberculomas who responded to treatment versus resistant cases.

Limitations:

The small sample size and retrospective analyses.

Ethics committee approval

Ethics committee approval obtained.

Funding:

No funding was received for this work.

Purpose:

Volumetric quantification of white matter hyperintensities (WMH) is well established in research and is starting to be adopted clinically for structured reports.

Automated tools using machine learning should avoid time-consuming manual segmentation and give more objective results. However, supervised tools require segmentation examples to train the algorithm and the tools’ performance is usually evaluated by comparing the output to manual masks, which suffer from between and within rater variability.

We aimed to evaluate whether our tool, FSL-BIANCA, can overcome the variability present in the manual training set and increase the consistency of automatic results.

Methods and materials:

We used 24 MRI scans from the Whitehall II imaging sub-study.

Manual WMH segmentation was performed by two raters (R1, R2) and repeated by the second rater a year later (R2a, R2b).

The manual masks were used to train BIANCA and the automated WMH masks were generated using a leave-one-out approach.

Between and within rater agreement on manually and automatically segmented masks were calculated using a Dice index and results were compared with paired t-tests.

Results:

The agreement between BIANCA outputs generated with masks from different raters is similar to the manual between rater agreement. BIANCA outputs trained with masks from R2 (a,b) are more consistent than the within rater agreement of manual masks.

Conclusion:

Our results suggest that if the examples provided to BIANCA are sufficiently in agreement, the automated tool improves the consistency of the output. This also highlights the need to standardise the definition of WMH, especially if automated tools are planned to be used in multicentre studies.

Limitations:

The limited amount of scans, raters (with different expertise), and re-tests.

Ethics committee approval

Ethics approval and consent from all participants were obtained.

Funding:

MRC (G1001354), MRC-DPUK, and Parkinson’s UK.

Purpose:

The automated brain volumetry from MRI images is a modern tool to quantify brain anatomy in patients with conditions related to brain atrophy. There is much software available in the market. We choose Neuroquant (NQ) and Volbrain (VB) in order to investigate any differences between their volumetric measurements and how these differences affect their atrophy estimation.

Methods and materials:

26 patients with multiple sclerosis visited the MRI lab at Bioiatriki SA (Athens) with the query of brain volumetry. They scanned in GE Discovery 750 3.0T MR system and the exam protocol included a 3D T1 sequence without contrast enhancement. The parameters in 3DT1 were in accordance with NQ's directions. These same images were processed both with NQ and VB. Measurements from different brain structures were compared. The unpaired t-test was the statistical method used for data analysis. In patients with an atrophy result from NQ, the possible agreement with VB was checked.

Results:

Whole-brain, hippocampus, cerebellum, and white matter measurements had or had not quite significant differences (p>0.05).

Intracranial cavity measurements had very significant differences.

Amygdale, putamen, and thalamus had extremely significant differences with a deviation from the bibliography average of 167.7% for NQ and 17.7% for VB, 35.9% for NQ and 3.3% for VB, and 0.7% for NQ and 29.3% for VB, respectively.

Conclusion:

There are important differences in the measurements of some brain structures, with significant deviation from the bibliography average for both NQ and VB.

Although the initially unexpected differences, NQ and VB are both able to determine atrophy.

VB has smaller deviations from average values and, for the time, is free (10/d).

Limitations:

The 3DT1 parameters should be constant during the study.

Ethics committee approval

n/a

Funding:

No funding was received for this work.

Purpose:

To evaluate the performance of deep learning AI (DLAI) to match routine lumbar spine MRI image quality at highly reduced scan times.

Methods and materials:

27 consecutive patients (49+/-16 years old; 17 male) underwent standard of care (SOC) lumbar spine MRI exams on 1 of 3 different clinical 1.5T scanners. All subjects underwent an additional accelerated 2D sagittal T2 series processed by an FDA cleared convolutional neural network-based deep learning AI application trained on multivendor MR platforms (SubtleMR™). The sagittal T2 scan times averaged 2:12 (SOC) and 0:49 (DLAI=2.7x acceleration). 54 image series (27 SOC and DLAI) were randomised and independently rated by two board-certified neuroradiologists for perceived SNR, anatomy/pathology conspicuity, motion artefacts, and overall image quality on a 5-point Likert scale (1: non-diagnostic, 2: poor, 3: diagnostic, 4: good, and 5: excellent). A two-sided paired t-test was performed with P<0.05 considered as statistically significant.

Results:

The average scores for perceived SNR, anatomy/pathology conspicuity, motion artefacts, and overall image quality (SOC/DLAI) were 5.0/4.9, 5.0/4.9, 4.9/4.8, and 5.0/4.9, respectively, for reader 1, and 4.4/4.3, 4.8/4.5, 4.7/4.7, and 4.4/4.2 for reader 2. No statistically significant difference between DL-accelerated scans and standard scans were present for all criteria and both readers.

Conclusion:

Deep learning AI technology can match routine lumbar spine MR image quality at approximately 1/3 of the scan time.

Limitations:

The limited number of subjects (27).

Ethics committee approval

IRB approved with written informed consent.

Funding:

No funding was received for this work.

Purpose:

Enlarged perivascular spaces (PVS) are a neuroimaging marker for cerebral small vessel disease. Manual annotation of PVS is challenging, time-consuming, and subject to observer bias. We developed and evaluated an automated method for the localisation of PVS in the two most clinically relevant regions: the centrum semiovale (CSO) and the basal ganglia (BG).

Methods and materials:

We used T2-weighted 1.5T MRI scans from 2,202 subjects enrolled in the population-based Rotterdam scan study, 1,202 for method development, and 1,000 for evaluation. An expert rater annotated one predefined slice per brain region with a single dot per PVS. A separate set of 40 MRI scans was annotated twice to estimate intra-rater agreement, which was computed by considering both sets once as ground truth and computing the average sensitivity and false-positives per image (FPPI).

We trained a regression convolutional neural network to predict the intensity-based distance to the nearest PVS for each voxel. Subsequently, the predicted distance maps were thresholded to detect PVS.

Results:

The intra-rater agreement showed an average sensitivity of 55.7% with 4.43 FPPI in the CSO and in the BG this was 73.2% and 2.09, respectively. Our method had similar performance with, for the same number of FPPI, an average sensitivity of 55.6% in the CSO and 75.3% in the BG.

Conclusion:

Even for an expert rater, this is a challenging problem as shown by the intra-rater agreement. Our proposed method can annotate PVS fully automatically as well as an expert rater. This method could replace the time-consuming manual assessment of PVS in neurological studies.

Limitations:

The single scanner, single protocol dataset; 1 slice per brain region annotated.

Ethics committee approval

Rotterdam scan study approved by the medical ethics committee of Erasmus MC.

Funding:

NWO-P15-26, Quantib, ZonMw104003005.

Purpose:

To develop a deep learning model (DLM) for fully automated detection and segmentation of intracranial aneurysm in patients with subarachnoid haemorrhages (SAH) on CT-angiography (CTA).

Methods and materials:

This retrospective single-centre study included 295 confirmed aneurysms from 253 patients with SAH (2010-2017). All patients from 2016-2017 (68 patients/79 aneurysms) provided the training-set, whereas subjects from 2010-2016 (185 patients/216 aneurysms) served as a test-set. Ground truth was established by independent manual segmentations of the aneurysms by a radiologist and neurosurgeon in a voxel-wise manner. CTA source images acquired using a standard clinical protocol for head/neck (n=222) and head (n=29) were pre-processed by extracting the brain mask and enhancing the blood vasculature. A 3D convolutional neural network (CNN) architecture based on DeepMedic was used for training 3 different DLMs using a different combination of vessel-enhanced input images as multiple-channels, as well as modifying the CNN parameters. An ensembling strategy was then used to generate the final DLM, optimised by using 5-fold-cross-validation.

Results:

In the training-set, DLM achieved an aneurysm detection rate of 0.72, average false-positives (FP’s)/scan of 0.21, a Dice coefficient of 0.74, and a volume correlation (r) of 0.97. Comparing the performance of the individual DLMs to the ensemble approach, we observed an increase of 37% (0.54 vs 0.74) with respect to Dice and a decrease of 90% with respect to FP’s/scan (2.10 vs 0.21). On the independent test-set, sensitivity was 0.82 with an average of 0.81 FP’s/scan, a Dice of 0.75, and r of 0.9.

Conclusion:

The proposed DLM is robust in detecting and segmenting aneurysms in patients with SAH on routine CTA images. We also demonstrated a significant increase of performance by using an ensemble of different DLMs.

Limitations:

n/a

Ethics committee approval

The local ethics committee approved this retrospective study.

Funding:

No funding was received for this work.

Purpose:

To evaluate the reproducibility of morphometry MRI to assess atrophy in patients with a mild cognitive impairment (MCI).

Methods and materials:

We retrospectively analysed MR images of 114 patients with MCI. One radiologist rated atrophy in 6 key brain regions bilaterally in T1-MPRAGE sequences using established visual rating scales.

Afterwards, atrophy was rated with a morphometry function (MorphoBox algorithm, Siemens prototype) which automatically estimates the volume of different brain regions and compares them with normative adjusted ranges. Maps with the different brain regions colour-coded indicated the degree of deviation (no deviations=green, mild=yellow, moderate=orange, and large=red).

Using Fisher’s tests, we evaluated the association between classifications (visual and morphometric) and between each classification and the actual clinical diagnosis, including all patients and including only those classified as normal or MCI/Alzheimer’s disease (AD) by all methods.

Results:

When all patients were included, there was no association between visual and morphometric classifications (p=0.44). However, if only normal or MCI/AD patients were considered, the association was close to significance (p=0.073).

When all patients were included there was a significant association between the visual classification and clinical diagnosis (p=0.006), but no association between morphometry and clinical diagnosis (p=0.29).

However, when only normal or MCI/AD patients were considered, the association between clinical diagnosis and visual classification was not significant (p=0.43), while the association between clinical diagnosis and morphometry was closer to significance (p=0.0920).

Conclusion:

The morphometry function has a tendency to be significant when it comes to differentiating normal patients from MCI/AD. Nevertheless, it doesn’t classify other dementias as well as visual classification.

Morphometry may represent a facilitating tool in the clinical assessment of MCI/AD, although further investigations are required.

Limitations:

The small sample size.

Ethics committee approval

Ethics committee approval obtained.

Funding:

No funding was received for this work.