RPS 1010a - Artificial intelligence (AI) and new techniques in MRI

Purpose:

To determine the value of MR texture analysis to assess tumour margin infiltration in soft tissue sarcomas at 3T MRI including IVIM DWI.

Methods and materials:

31 patients with pathologically confirmed soft tissue sarcomas after surgery who underwent 3T MRI including IVIM DWI were included in this study. Margin infiltration on pathology was used as a golden standard. Texture analysis was performed by drawing ROIs on axial T1- and T2-weighted images, contrast-enhanced (CE) T1-weighted images, high b value DWI, and ADC maps using software (TexRad). Mean, SD, entropy, MPP, skewness, and kurtosis were compared between sarcomas with an infiltrative margin and sarcomas with a circumscribed margin. A Mann-Whitney U test was performed. The ROC curve with AUC was obtained.

Results:

Among 31 patients with sarcomas, 22 sarcomas had tumour margin infiltration at pathology. T1 kurtosis was significantly higher in sarcomas with an infiltrative margin than a circumscribed margin: 1.959 vs 0.12 (p=0.017). CE T1 kurtosis was significantly lower in sarcomas with an infiltrative margin than a circumscribed margin: 0.563 vs 1.413 (p=0.020). The mean and MPP on DWI were higher in sarcomas with an infiltrative margin than a circumscribed margin: 404.4 vs 269.4 (p=0.082). AUCs of T1 kurtosis and skewness were 0.792 (95% CI, 0.628-955) and 0.601 (95% CI, 0.371-830), respectively. On DWI, AUC of mean and MPP was 0.702 (95% CI, 0.520-0.884). AUC of CE T1 mean and MPP was 0.630 (95% CI, 0.383-877). With a cutoff of 0.545 in T1 kurtosis, sensitivity and specificity for predicting tumour margin infiltration were 71% and 88%, respectively.

Conclusion:

MR texture analysis may be reliable to predict the tumour margin infiltration in soft tissue sarcomas, particularly using T1 kurtosis at 3T.

Limitations:

A retrospective study with potential selection bias and a small study population.

Ethics committee approval

Approved by IRB and informed consent was waived.

Funding:

No funding was received for this work.

Purpose:

Hip dysplasia and FAI are complex three-dimensional hip pathologies that can cause hip pain in young and active patients of child-bearing age. Imaging is static and based on 2D pelvic radiographs. Computer-assisted CT-based diagnosis of hip dysplasia and FAI was introduced for patient-specific planning of surgical treatment but MRI-based 3D models would offer a radiation-free alternative.

This study explores the difference between automatic and manual MRI-based 3D-models and if diagnostic parameters correlate between automatic and manual MR-based 3D-models.

Methods and materials:

We performed an IRB-approved comparative, retrospective study of 31 hips of 26 symptomatic patients with FAI and hip dysplasia. We compared CT-and MR-based osseous 3D models of the hip joint of the same patients. A 3D-CT scan (isovoxel: 1 mm³) of the entire pelvis and the distal femoral condyles were obtained. Preoperative MR arthrograms of the hip were obtained including axial-oblique T1 3D-VIBE sequence (0.8 mm³ isovoxel) and two axial anisotropic (1.2 x1.2 x 1 mm) 3D-T1 VIBE DIXON sequences of the entire pelvis and the distal femoral condyles. Automatic reconstruction of MRI-based 3D models was performed using machine-learning (deep learning). Threshold-based manual reconstruction of CT- and MR-based 3D-models was performed using commercial software (AMIRA).

Results:

The difference between MRI-based 3D-models was below 1 mm. The median difference was 0.2±0.1 mm for the proximal femur and 0.2±0.5 mm for the acetabulum. The Dice coefficient was 97% for the acetabulum and 98% for the femur.
The correlation for 6 diagnostic parameters was excellent and significant (r=0.99, p<0.001) between automatic and manual MR-based 3D-models. The absolute difference was below 2°.

Conclusion:

Automatic MRI-based 3D models can replace CT-based 3D-models for patients of childbearing age with hip dysplasia and FAI.

Limitations:

A retrospective study.

Ethics committee approval

IRB approval was obtained.

Funding:

No funding was received for this work.

Purpose:

Ultrasound evaluation of superficial soft tissue masses is increasingly utilised in routine clinical practice. However, differentiating benign from malignant aetiologies may be challenging. Deep convolutional neural networks (CNNs) have shown the potential to classify images with good accuracy. We propose to test the ability of this methodology to differentiate between benign and malignant superficial soft tissue masses.

Methods and materials:

A total of 760 ultrasound images from focused ultrasounds of superficial soft tissue masses were selected to train a CNN classifier. The dataset consisted of 523 images of benign soft tissue masses and 237 images of malignant masses, all confirmed by pathologic diagnosis. We selected the pre-trained VGG-16 architecture implemented on Keras. We performed a binary classification (benign and malignant). Then, we tried to differentiate between 3 subgroups of benign masses: lipomas, benign nerve sheath tumours, and vascular malformations.

Results:

The test accuracy of the model to determine if the mass was benign or malignant mass was 0.75. The CNN outperformed two experienced musculoskeletal radiologists on a test sample dataset. The accuracy of the model to differentiate between the 3 benign subgroups test dataset was 0.66.

Conclusion:

As an initial step, artificial intelligence (AI) shows promise in the differentiation of benign from malignant soft masses, outperforming two musculoskeletal radiologists. The addition of a larger variety and number of soft tissue masses will be necessary to establish the utility of AI as a useful tool going forward.

Limitations:

The limited size of our dataset restricted the number of subgroups that could be investigated. We tested the accuracy of only two radiologists, which does not necessarily reflect the accuracy of radiologists with varying experience.

Ethics committee approval

This study has institutional board review approval.

Funding:

No funding was received for this work.

Purpose:

Tibial fracture detection is a routine task in paediatric radiology. Its precision depends on knowledge of anatomy and experience-based training. Until recently, there were no computer-aided detection tools due to anatomical complexity, a variety of 2d projections, additional objects in the image, the presentation of pathology, the age of patients, and gender.

Methods and materials:

We presented a CNN-based system that can automatise radiographic distal tibia fracture detection. The system used two connected CNNs. The first CNN was trained to localise the distal tibia using 500 normal and 500 pathological studies, including displaced and nondisplaced fractures. The images were manually annotated with a box containing the distal tibia. This CNN had a 98% accuracy with zero false-positives. The second CNN was chosen from previously developed CNNs trained on a different, broader set of manually segmented distal tibia images and had undergone the 10 fold cross-validation. The best CNN, which formerly achieved an accuracy of 95.9%, was selected for validation in this study.

Results:

The whole system combined two models into the pipeline. An x-ray image was read and the distal tibia localised as an ROI. The ROI was transferred to the fracture detection algorithm. If the distal tibia was not localised, then it was not processed by the second model. The whole system achieved an area under the receiver-operating curve (AUC-ROC) of 0.89.

Conclusion:

This method could become a computer-aided diagnosis (CAD) tool, accelerating the workflow in paediatric radiology departments.

Limitations:

This study was limited to images acquired and read at a single institution.

Ethics committee approval

This work was conducted under a protocol approved by the Institutional Review Board.

Funding:

No funding was received for this work.

Purpose:

To test our hypothesis that three-dimensional phenotyping of the lumbosacral spine and dural sac using shape-based machine learning enables the prediction of Fibrillin-1 gene mutations pathogenic for Marfan syndrome.

Methods and materials:

184 patients being evaluated for Marfan syndrome, 01/2012-12/2016, underwent 1.5T-MRI with 3D T2-weighted TSE-sequence of the lumbosacral spine with 1x1x1 mm³ spatial resolution. 110 patients (32.4±11.6 years, 50 female) agreed to genetic testing; 38 had a Fibrillin-1 gene mutation pathogenic for Marfan syndrome (FBN1+) and 72 were tested negative (FBN1-). Shape-based machine learning allowed three-dimensional segmentation and quantification of volumes and volume ratios of vertebral bodies L3-L5 and dural sac segments L3-S1.

Results:

Dural sac volumes were significantly enlarged in FBN1+ vs FBN1- patients (in mL): segment L3 12.5±3.0 vs 10.3±2.3, L4 11.8±3.5 vs 9.0±2.5, L5 11.7±4.4 vs 7.9±3.1, and S1 13.5±8.7 vs 5.3±3.1 (all p<0.001). Vertebral body volumes did not significantly differ between FBN1+ vs FBN1- patients. ROC analysis for the identification of FBN1+ patients showed AUCs of 0.797 for total lumbosacral dural sac volume, 0.660 for L3 volume ratio, 0.694 for L4 volume ratio, 0.766 for L5 volume ratio, 0.786 for S1 dural sac volume, and 0.753 for S1 to L4 dural sac volume ratio. Combined sensitivity, specificity, and positive and negative predictive values for identification of FBN1+ patients were 69%, 93%, 84%, and 85%.

Conclusion:

Fibrillin-1 gene mutations were associated with a significantly altered three-dimensional phenotype of the lumbosacral spine and dural sac. Shape-based machine learning allowed identification of patients with Fibrillin-1 gene mutations with a high specificity and moderate sensitivity.

Limitations:

Potential sources of bias are the retrospective study design and secondary exclusion of those patients without genetic testing.

Ethics committee approval

An ethics committee approved the study and waived written informed consent.

Funding:

No funding was received for this work.

Purpose:

Post-mortem identification of unknown deceased persons and age estimation in the living are central tasks of forensic scientists. Body height is traditionally derived from measurements of long bones. The sacrum is a centrally located bone that is relatively unsusceptible to environmental factors. The aim of the study was to investigate if sacrum height can be used to estimate body height, weight, age, and sex using datasets from a population-based magnetic resonance imaging (MRI) study.

Methods and materials:

A total of 2,499 whole-body MRIs were evaluated. Sacrum height was measured in the median plane of sagittal T2-weighted images of the spine and associations with age, sex, and body height were analysed. Linear regression was used to derive formulas for body height estimation.

Results:

Male participants had significantly (p<0.001) greater sacrum height (11.4 cm, standard deviation 1.1 cm, range: 7.9-14.6 cm) compared to female participants (10.9 cm, standard deviation 1.0 cm, range: 7.3-14.5 cm). The Pearson-correlation was r=0.44 for body height. The mean absolute difference (MAD) between calculated and measured body height was 4.9 cm for females and 4.8 cm for males. ROC cut-off to classify sex had a 60% sensitivity and 57.3% specificity.

Conclusion:

Contrary to previous reports, sacrum height alone cannot be recommended to predict body height, age, or sex. It is not as reliable as predictions based on large bones.

Limitations:

A population-based MR-study from a specific region of Germany. Not all participants in the study were included in the final analysis. There was no correlation with established techniques for age, sex, and height determination based on the analysis of long bones.

Ethics committee approval

Approved by the local ethics committee and all participants gave their written informed consent.

Funding:

No funding was received for this work.

Purpose:

To describe the most common locations of involvement in chronic recurrent multifocal osteomyelitis (CRMO) in children using whole-body magnetic resonance imaging (WB-MRI).

Methods and materials:

We retrospectively reviewed 29 patients (12 girls, 17 boys) with a diagnosis of CRMO from 2014-May 2019. All patients underwent WB-MRI using STIR sequences in axial and coronal planes, and sagittal planes to evaluate the spine. Imaging data was evaluated by a paediatric radiologist with more than 10 years of experience and a radiology fellow with 3 years of experience.

Results:

The mean age at diagnosis was 11 years (3-16 years), but the mean age of clinical presentation was 10 years.

A total of 175 lesions were investigated. The lesions were multifocal in 28/29 cases (96.5%) and the number of lesions encountered per patient ranged from 1-28 lesions.

The most common sites of involvement were the pelvis and femur (21%), followed by tibia (19.3%), and the spine and humerus (7.4%). Other less frequent locations included clavicles (4.5%), tarsus (4%), foot (3.4%), radius and cuboid (2.8%), fibula (2.3%), and the sternum (1.7%). The involvement of the patella, ribs, carpus, and scapula was exceedingly rare, accounting for 0.6%.

The metaphysis of long bones was most frequently involved (56.4%), followed by epiphysis (13.6%). Bilateral lesions were 42.9%.

Conclusion:

WB-MRI is the diagnostic modality of choice in patients with a clinical diagnosis of CRMO. Bilateral involvement of metaphysis and epiphysis of long bones, and the pelvis, are the most frequent areas of involvement.

Limitations:

n/a

Ethics committee approval

n/a

Funding:

No funding was received for this work.

Purpose:

To evaluate the clinical performance of deep learning (DL)-based software for bone age assessment in a paediatric endocrinology clinic in South Korea.

Methods and materials:

We compared the bone age estimated from commercial DL-based software (BoneAge, Vuno, Seoul, Korea) with the bone ages independently estimated by 3 physicians (1 musculoskeletal radiologist, 1 paediatric endocrinologist, and 1 radiology resident) using a Greulich-Pyle atlas. In total, 109 children (8 boys, 101 girls) aged between 4 and 17 years who visited the paediatric endocrinology clinic from December 2018 were enrolled. Fisher’s exact test (two sides), Pearson’s correlation coefficient, and root mean squared error (RMSE) were used to compare the bone ages of DL software and those of the 3 physicians. An intraclass correlation coefficient (ICC) evaluated inter-rater variation.

Results:

Fisher’s exact test showed significant differences between DL bone age and all 3 physicians’ bone age (p<0.025). There were good correlations (r=0.93, 0.92, and 0.9, p<0.05), however, the RMSE values were 9.7, 13.1, and 9.9 months between the bone age of DL software and those of the musculoskeletal radiologist, paediatric endocrinologist, and radiology resident, respectively. RMSE values between the 3 physicians were 12.3, 18, and 14.8 months, with relatively good correlation with each other (r=0.88, 0.89, and 0.88, p<0.05). ICC values between the 3 physicians were 0.93, 0.84, and 0.87 each, with an overall ICC value of 0.92.

Conclusion:

Bone age estimated by DL-based software showed good correlation with the bone ages estimated using a traditional Greulich-Pyle atlas by 3 physicians but showed statistically different values.

Limitations:

A retrospective, single-centre study with a small number of patients

Ethics committee approval

n/a

Funding:

No funding was received for this work.