Research Presentation Session

RPS 1005a - Radiomics and texture analysis

Lectures

1
RPS 1005a - Differentiating head and neck paraganglioma versus schwannoma using texture analysis: a preliminary analysis

RPS 1005a - Differentiating head and neck paraganglioma versus schwannoma using texture analysis: a preliminary analysis

08:15S. Malla, New Delhi / IN

Purpose:

To investigate the value of T2-weighted based radiomics in the differentiation of head and neck paragangliomas and schwannomas.

Methods and materials:

This retrospective study included patients presenting with neck masses who were imaged under our institutional protocol in 3T MRI scanner between January 2016 and March 2019. A total of 46 patients (36 paragangliomas and 16 schwannomas) were identified using a composite gold standard of histopathology, cytology, and DOTANOC PET CT. Fat-suppressed T2-DIXON axial images with TR-2500-3500msec, TE-90msec, 4 mm slice thickness and 190x190 mm-FOV were used for analysis.

First- and second-order texture-parameters were calculated from original and filtered images. Feature selection using F-statistics and collinearity analysis provided 10 texture parameters for further analysis. A Mann-Whitney-U test was used to compare the two groups and p-values were adjusted for multiple comparisons. AUC for the significant features was obtained.

Results:

A total of 7 texture parameters were found to be significantly different between paragangliomas and schwannomas (adjusted p < .05). Laplacian of Gaussian intermediate, first-order kurtosis, wavelet HLL, grey-level co-occurrence matrix (glcm) correlation, and wavelet LHH glcm correlation each achieved a specificity of 93.5% in differentiating paraganglioma and schwannoma with AUC of 0.738, 0.819, and 0.689, respectively. A logistic regression model combined the 7 significant texture parameters to differentiate between paragangliomas and schwannomas, which obtained a sensitivity of 97.2% and specificity of 75%, with an AUC of 0.97 (0.87 to 0.99).

Conclusion:

T2-weighted based radiomics could serve as a useful adjunct to conventional MRI in distinguishing head and neck paragangliomas from schwannomas, which may influence triage contrast administration or nuclear medicine based DOTANOC studies.

Limitations:

These results are preliminary and thus require validation on a larger and independent data set to assess reproducibility for the potential for clinical translation.

Ethics committee approval

n/a

Funding:

No funding was received for this study.

2
RPS 1005a - Radiomics-based approach for the diagnosis of osteoporosis using hip radiographs

RPS 1005a - Radiomics-based approach for the diagnosis of osteoporosis using hip radiographs

06:05Kim Sang Wook, Seoul / KR

Purpose:

To evaluate the performance of a radiomics-model for the diagnosis of osteoporosis using hip radiographs.

Methods and materials:

For the development dataset, 6,995 hip anteroposterior radiographs with dual-energy x-ray absorptiometry (DXA) were collected between 2010 and 2018. Both hip joints were automatically segmented using U-net, except for joints with fractures or metal prostheses. A total of 11,243 joints (normal 9,170, osteoporosis 2,073) were used for the development dataset and a temporally-separated validation dataset was comprised of 553 joints (normal 459, osteoporosis 74) from 300 hip anteroposterior radiographs between 2008 and 2009. T-scores on hip DXA were used as reference standards. Three clinical (age, sex, and weight) and 293 radiomics features were used to train the random forest and gradient boosting classifiers. Two radiologists assessed the possibility of osteoporosis using a 5-point scale and the performance of the radiologist was compared with the radiomics-model. Receiver-operating characteristic curve analysis was used to assess diagnostic performance.

Results:

The diagnostic performance was AUC 0.809 (95% CI 0.774-0.842) for board-certified radiologists and AUC 0.760 (95% CI 0.721-0.796) for the radiology resident. The performance of the radiomics-model was AUC 0.830 (95% CI 0.796-0.861), sensitivity 74.3%, and specificity 79.5%, with an optimal cut-off probability of 0.318. The performance of the radiomics-model was significantly higher than the radiology resident (P = 0.038).

Conclusion:

The radiomics-model for the diagnosis of osteoporosis using hip radiographs showed comparable diagnostic performance to those of the radiologists and may help in the automatic screening of osteoporosis in plain radiographs.

Limitations:

Additional external validation is needed to assess the generalisability of the model.

Ethics committee approval

All datasets were collected under IRB approval in SNUH.

Funding:

This work was supported by the National Research Foundation of Korea grant funded by the Korean government (No. NRF-2017M2A2A6A01071214).

3
RPS 1005a - Can MRI texture analysis of the pancreas predict postoperative pancreatic fistulas?

RPS 1005a - Can MRI texture analysis of the pancreas predict postoperative pancreatic fistulas?

05:44S. Skawran, Zürich / CH

Purpose:

To evaluate whether a magnetic resonance imaging (MRI) texture analysis (TA)-based machine learning (ML) classifier can predict postoperative pancreatic fistulas (POPF) after pancreaticoduodenectomy (PD).

Methods and materials:

70 patients who underwent MRI before PD between 2010 and 2018 were retrospectively analysed. POPF was graded according to the international study group of pancreatic fistula and split into clinically relevant versus non-relevant or no POPF. On T1- and T2-weighted images, two regions of interest were placed in the pancreatic corpus and cauda. 1,350 radiomic features were extracted after standardised image processing using the open-source pyRadiomics library. The dataset was split into training and test sets. After feature reduction, a non-linear support vector machine (SVM) was trained on the 3 most important features. Logistic regression analyses were performed using these features and mean T1 and T2 signal intensity values (SImean). Diagnostic accuracy of the models was compared using areas under the receiver operating characteristics curve (AUCs).

Results:

One T1 and two T2 higher-order texture features were identified as the most important variables for classification. All 3 variables showed significant differences between the two groups (all p<0.02). The trained SVM achieved an AUC of 0.68 when applied on the test dataset. In contrast, logistic regression models using only SImean from T1- and T2-weighted images resulted in an AUC of 0.51 [T1mean], 0.60 [T2mean], and 0.59 [combined] (all p<0.001).

Conclusion:

MRI-texture analysis based on routine sequences provides a promising prediction of clinically relevant POPF.

Limitations:

Studies were limited to one vendor. Cross-vendor-transferability remains to be elucidated. Image analysis was done by one reader. Reproducibility will be subject to further research.

Ethics committee approval

The local ethics committee approved this retrospective study and all patients gave written informed consent.

Funding:

No funding was received for this work.

4
RPS 1005a - The influence of different levels of adaptive statistical iterative reconstruction (ASIR) regarding computed tomography texture features

RPS 1005a - The influence of different levels of adaptive statistical iterative reconstruction (ASIR) regarding computed tomography texture features

05:25T. Polidori, Roma / IT

Purpose:

To evaluate the influence of different levels of ASIR on computed tomography texture features.

Methods and materials:

We analysed 10 patients who underwent unenhanced computed tomography (CT) scans of the abdomen with the same scanner (Revolution Evo, GE Healthcare, USA). Subsequently, we reconstructed raw data with 11 levels of ASIR (from 0 to 10), thus obtaining datasets with different percentages of blending between filtered back projection and iterative reconstruction. Two radiologists analysed the texture features of the liver, kidney, spleen, and muscle tissues using four different regions of interest (ROIs) that were cloned for all 11 different iteration levels datasets. Data was then elaborated with TexRad medical imaging software. 5 different radiomic features (mean, standard deviation, entropy, skewness, and kurtosis) were compared among the different reconstruction algorithms and spatial scaling factor (SSF) by using ANOVA test with Bonferroni correction.

Results:

Different iteration levels significantly affect (p<0.05) standard deviation (SD) and entropy values for SSF 0, with a decrease of these parameters from 0 to 10 ASIR (e.g. statistically significant decrease of entropy between level 10 and 0-7). The influence of different ASIR levels on these parameters gradually faded with an increase in the SSF, thus no significant differences between all the radiomic values have been obtained from SSF 5 to 6. We did not find statistically significant differences between the values of mean, skewness, and kurtosis, regardless of the SSF value and the tissue.

Conclusion:

We demonstrated that different ASIR levels moderately affect standard deviation and entropy with a declining trend from SSF 0 to SSF 5-6. Mean, skewness, and kurtosis are not significantly affected by different ASIR levels.

Limitations:

A small study population.

Ethics committee approval

n/a

Funding:

No funding was received for this work.

5
RPS 1005a - Supervised machine learning predictive modelling for primary versus secondary lung malignancy in CT-guided transthoracic biopsies

RPS 1005a - Supervised machine learning predictive modelling for primary versus secondary lung malignancy in CT-guided transthoracic biopsies

05:57E. Barbosa Jr., Philadelphia / US

Purpose:

CT-guided transthoracic biopsy (CTTB) is a minimally invasive method for the diagnostic evaluation of a variety of thoracic diseases. We leveraged a large cohort of CTTB patients to assess how well supervised machine learning models utilising clinical and imaging variables can predict pathology diagnosis in primary versus secondary lung malignancy.

Methods and materials:

588 patients were retrospectively identified, comprising of biopsies performed at our institution. Our model included variables such as age, race, sex, smoking history, smoking pack-years, history of prior cancer, immune status, nodule location, nodule size, nodule margin, and nodule shape. These variables were extracted from radiology reports and electronic medical records. Ground-truth was CTTB pathology diagnosis, classified as primary (60.5%) versus secondary (39.5%). We compared 3 supervised machine learning classifiers: random forest (RF), logistic regression (LR), and decision tree (DT), utilising a 75%/25% training/validation ratio for the CTTB cohort (10-fold cross-validation).

Results:

The RF model produced AUC (area under the curve) of ROC (receiver operating characteristic) of 0.87 and accuracy of 0.81, the LR model produced AUC of ROC of 0.89 and accuracy of 0.81, whereas the DT model produced AUC of ROC of 0.79 and accuracy of 0.71.

Conclusion:

CTTB, while accurate for pathologic diagnosis of thoracic lesions, can be costly and may entail significant risks. LR and RF models achieved up to 81% diagnostic accuracy, potentially allowing non-invasive prediction of primary versus secondary lung malignancy, a crucial distinction for patient management.

Limitations:

A retrospective design.

Ethics committee approval

IRB approved, waiver of informed consent.

Funding:

Funded by the NIH (USA).

6
RPS 1005a - Dominant language hemisphere detection by using a ML model informed by structural and functional connectivity

RPS 1005a - Dominant language hemisphere detection by using a ML model informed by structural and functional connectivity

05:44J. PARIENTE, BARCELONA / ES

Purpose:

To develop a method to determine the language dominant hemisphere (LDH) by using structural, diffusion, and resting-state MRI.

Methods and materials:

This study was performed on structural and functional neuroimaging data from 15 subjects. We define the LDH goal standard using MRI functional data with language and comprehension tasks.

For the structural analyses, whole-brain tractography reconstructions were performed using multi-shell, multi-tissue, constrained spherical deconvolution. We obtained the language-related tracts using a pre-trained convolutional neural network (Tractseg) and used them to obtain language connectivity matrices.

We obtained different indices from the volumetric acquisition (e.g. cortical thickness), diffusion-based parameters (e.g. FA), and graph theoretical indices derived from structural and functional language connectivity matrices. We performed a parametric statistical analysis for each of the metrics and used the significant indices to obtain single-subject laterality differences by training a ML-based classification algorithm.

Additionally, we compared the results obtained with our methodology using the standard laterality index approach.

Results:

We found statistically significant differences for the cortical thickness in the superior temporal gyrus and in the MD for the uncinated tract. The ML model had an accuracy of 0.90 +/- 0.19 with a confusion matrix.

We did not find differences using the LI analysis.

Conclusion:

Our results suggest that we can use this methodology to predict the LDH by using a ML classification model based on resting-state and diffusion from MRI data.

Limitations:

The sample is not big enough for a ML-based study. This abstract is a proof of concept and we are now acquiring new subjects to validate the new methodology.

Ethics committee approval

n/a

Funding:

This study is funded by a FIS project PI/012 and the European funding of regional development (FEDER).

7
RPS 1005a - Quantitative ultrasound texture analysis of the foetal lung versus foetal pulmonary artery Doppler as a non-invasive predictor of neonatal respiratory distress syndrome (RDS)

RPS 1005a - Quantitative ultrasound texture analysis of the foetal lung versus foetal pulmonary artery Doppler as a non-invasive predictor of neonatal respiratory distress syndrome (RDS)

10:36N. Osman, Minia / EG

Purpose:

To make a comparison between quantitative ultrasound texture analysis of the foetal lung (QuantusFLM) and pulmonary artery acceleration time/ejection time ratio (AT/ET) for the evaluation of foetal lung maturity.

Methods and materials:

The prospective cohort study was conducted on 65 pregnant women selected from cases referred to the obstetrics and gynaecology department at Minia University Hospital from May 2018 to March 2019. All patients underwent ultrasound examination for foetal pulmonary artery acceleration time/ejection time ratio (AT/ET) as well as a grey-scale axial ultrasound image of the foetal lungs, which were analysed using the Quantus-FLM software online application. 25 patients were excluded from the study due to a foetal birth more than 48 hours after the ultrasound examination, corticosteroid administration in between the examination and delivery, congenitally malformed foetuses, neonatal sepsis, or respiratory distress due to causes other than neonatal RDS. After delivery, the neonates were grouped according to diagnosis of RDS as positive or negative.

Results:

From the 40 eligible foetuses, 9 (22%) developed neonatal RDS. Both AT/ET and Quantus FLM results were significantly correlated with the development of neonatal RDS. AT/ET were significantly lower in feotuses developed RDS. A cutoff value of 0.305 for AT/ET predicted the development of RDS (sensitivity: 77.78%, specificity: 87.1%) while Quantus FLM predicted neonatal RDS with 88.9% sensitivity and 90.32% specificity.

Conclusion:

Both AT/ET and Quantus-FLM may provide a noninvasive means of detecting the development of neonatal RDS with acceptable levels of sensitivity and specificity.

Limitations:

Small numbers of patient study.

Ethics committee approval

This study is approved by our university ethics committee.

Funding:

No funding was received for this work.

8
RPS 1005a - Classification of non-enhancing glioma tumours by diffusion MRI with B-tensor encoding: an explorative study of tumour heterogeneity

RPS 1005a - Classification of non-enhancing glioma tumours by diffusion MRI with B-tensor encoding: an explorative study of tumour heterogeneity

05:43J. Brabec, Lund / SE

PEP Subscription Required

This course is only accessible for ESR Premium Education Package subscribers.