Image annotation and curation in radiology: an overview for machine learning practitioners

"Garbage in, garbage out" summarises well the importance of high-quality data in machine learning and artificial intelligence. All data used to train and validate models should indeed be consistent, standardised, traceable, correctly annotated, and de-identified, considering local regulations. This narrative review presents a summary of the techniques that are used to ensure that all these requirements are fulfilled, with special emphasis on radiological imaging and freely available software solutions that can be directly employed by the interested researcher. Topics discussed include key imaging concepts, such as image resolution and pixel depth; file formats for medical image data storage; free software solutions for medical image processing; anonymisation and pseudonymisation to protect patient privacy, including compliance with regulations such as the Regulation (EU) 2016/679 "General Data Protection Regulation" (GDPR) and the 1996 United States Act of Congress "Health Insurance Portability and Accountability Act" (HIPAA); methods to eliminate patient-identifying features within images, like facial structures; free and commercial tools for image annotation; and techniques for data harmonisation and normalisation.Relevance statement This review provides an overview of the methods and tools that can be used to ensure high-quality data for machine learning and artificial intelligence applications in radiology.Key points• High-quality datasets are essential for reliable artificial intelligence algorithms in medical imaging.• Software tools like ImageJ and 3D Slicer aid in processing medical images for AI research.• Anonymisation techniques protect patient privacy during dataset preparation.• Machine learning models can accelerate image annotation, enhancing efficiency and accuracy.• Data curation ensures dataset integrity, compliance, and quality for artificial intelligence development.

Normal cohorts in automated brain atrophy estimation: how many healthy subjects to include?

OBJECTIVES

This study investigates the influence of normal cohort (NC) size and the impact of different NCs on automated MRI-based brain atrophy estimation.

METHODS

A pooled NC of 3945 subjects (NCpool) was retrospectively created from five publicly available cohorts. Voxel-wise gray matter volume atrophy maps were calculated for 48 Alzheimer's disease (AD) patients (55-82 years) using veganbagel and dynamic normal templates with an increasing number of healthy subjects randomly drawn from NCpool (initially three, and finally 100 subjects). Over 100 repeats of the process, the mean over a voxel-wise standard deviation of gray matter z-scores was established and plotted against the number of subjects in the templates. The knee point of these curves was defined as the minimum number of subjects required for consistent brain atrophy estimation. Atrophy maps were calculated using each NC for AD patients and matched healthy controls (HC). Two readers rated the extent of mesiotemporal atrophy to discriminate AD/HC.

RESULTS

The maximum knee point was at 15 subjects. For 21 AD/21 HC, a sufficient number of subjects were available in each NC for validation. Readers agreed on the AD diagnosis in all cases (Kappa for the extent of atrophy, 0.98). No differences in diagnoses between NCs were observed (intraclass correlation coefficient, 0.91; Cochran's Q, p = 0.19).

CONCLUSION

At least 15 subjects should be included in age- and sex-specific normal templates for consistent brain atrophy estimation. In the study's context, qualitative interpretation of regional atrophy allows reliable AD diagnosis with a high inter-reader agreement, irrespective of the NC used.

CLINICAL RELEVANCE STATEMENT

The influence of normal cohorts (NCs) on automated brain atrophy estimation, typically comparing individual scans to NCs, remains largely unexplored. Our study establishes the minimum number of NC-subjects needed and demonstrates minimal impact of different NCs on regional atrophy estimation.

KEY POINTS

• Software-based brain atrophy estimation often relies on normal cohorts for comparisons. • At least 15 subjects must be included in an age- and sex-specific normal cohort. • Using different normal cohorts does not influence regional atrophy estimation.

Variation in cartilage T2 and T2* mapping of the wrist: a comparison between 3- and 7-T MRI

BACKGROUND

To analyze regional variations in T2 and T2* relaxation times in wrist joint cartilage and the triangular fibrocartilage complex (TFCC) at 3 and 7 T and to compare values between field strengths.

METHODS

Twenty-five healthy controls and 25 patients with chronic wrist pain were examined at 3 and 7 T on the same day using T2- and T2*-weighted sequences. Six different regions of interest (ROIs) were evaluated for cartilage and 3 ROIs were evaluated at the TFCC based on manual segmentation. Paired t-tests were used to compare T2 and T2* values between field strengths and between different ROIs. Spearman's rank correlation was calculated to assess correlations between T2 and T2* time values at 3 and 7 T.

RESULTS

T2 and T2* time values of the cartilage differed significantly between 3 and 7 T for all ROIs (p ≤ 0.045), with one exception: at the distal lunate, no significant differences in T2 values were observed between field strengths. T2* values differed significantly between 3 and 7 T for all ROIs of the TFCC (p ≤ 0.001). Spearman's rank correlation between 3 and 7 T ranged from 0.03 to 0.62 for T2 values and from 0.01 to 0.48 for T2* values. T2 and T2* values for cartilage varied across anatomic locations in healthy controls at both 3 and 7 T.

CONCLUSION

Quantitative results of T2 and T2* mapping at the wrist differ between field strengths, with poor correlation between 3 and 7 T. Local variations in cartilage T2 and T2* values are observed in healthy individuals.

RELEVANCE STATEMENT

T2 and T2* mapping are feasible for compositional imaging of the TFCC and the cartilage at the wrist at both 3 and 7 T, but the clinical interpretation remains challenging due to differences between field strengths and variations between anatomic locations.

KEY POINTS

•Field strength and anatomic locations influence T2 and T2* values at the wrist. •T2 and T2* values have a poor correlation between 3 and 7 T. •Local reference values are needed for each anatomic location for reliable interpretation.

Using automated software evaluation to improve the performance of breast radiographers in tomosynthesis screening

OBJECTIVE

To improve breast radiographers' individual performance by using automated software to assess the correctness of breast positioning and compression in tomosynthesis screening.

MATERIALS AND METHODS

In this retrospective longitudinal analysis of prospective cohorts, six breast radiographers with varying experience in the field were asked to use automated software to improve their performance in breast compression and positioning. The software tool automatically analyzes craniocaudal (CC) and mediolateral oblique (MLO) views for their positioning quality by scoring them according to PGMI classifications (perfect, good, moderate, inadequate) and checking whether the compression pressure is within the target range. The positioning and compression data from the studies acquired before the start of the project were used as individual baselines, while the data obtained after the training were used to test whether conscious use of the software could help the radiographers improve their performance. The percentage of views rated perfect or good and the percentage of views in target compression were used as overall metrics to assess changes in performance.

RESULTS

Following the use of the software, all radiographers significantly increased the percentage of images rated as perfect or good in both CCs and MLOs. Individual improvements ranged from 7 to 14% for CC and 10 to 16% for MLO views. Moreover, most radiographers exhibited improved compression performance in CCs, with improvements up to 16%.

CONCLUSION

Active use of a software tool to automatically assess the correctness of breast compression and positioning in breast cancer screening can improve the performance of radiographers.

CLINICAL RELEVANCE STATEMENT

This study suggests that the use of a software tool for automatically evaluating correctness of breast compression and positioning in breast cancer screening can improve the performance of radiographers on these metrics, which may ultimately lead to improved screening outcomes.

KEY POINTS

• Proper breast positioning and compression are critical in breast cancer screening to ensure accurate diagnosis. • Active use of the software increased the quality of craniocaudal and mediolateral oblique views acquired by all radiographers. • Improved performance of radiographers is expected to improve screening outcomes.

Complexities of deep learning-based undersampled MR image reconstruction

Artificial intelligence has opened a new path of innovation in magnetic resonance (MR) image reconstruction of undersampled k-space acquisitions. This review offers readers an analysis of the current deep learning-based MR image reconstruction methods. The literature in this field shows exponential growth, both in volume and complexity, as the capabilities of machine learning in solving inverse problems such as image reconstruction are explored. We review the latest developments, aiming to assist researchers and radiologists who are developing new methods or seeking to provide valuable feedback. We shed light on key concepts by exploring the technical intricacies of MR image reconstruction, highlighting the importance of raw datasets and the difficulty of evaluating diagnostic value using standard metrics.Relevance statement Increasingly complex algorithms output reconstructed images that are difficult to assess for robustness and diagnostic quality, necessitating high-quality datasets and collaboration with radiologists.Key points• Deep learning-based image reconstruction algorithms are increasing both in complexity and performance.• The evaluation of reconstructed images may mistake perceived image quality for diagnostic value.• Collaboration with radiologists is crucial for advancing deep learning technology.

Metal artifact reduction on musculoskeletal CT: a phantom and clinical study

BACKGROUND

Artifacts caused by metal implants are challenging when undertaking computed tomography (CT). Dedicated algorithms have shown promising results although with limitations. Tin filtration (Sn) in combination with high tube voltage also shows promise but with limitations. There is a need to examine these limitations in more detail. The purpose of this study was to investigate the impact of different metal artefact reduction (MAR) algorithms, tin filtration, and ultra-high-resolution (UHR) scanning, alone or in different combinations in both phantom and clinical settings.

METHODS

An ethically approved clinical and phantom study was conducted. A modified Catphan® phantom with titanium and stainless-steel inserts was scanned with six different MAR protocols with tube voltage ranging from 80 to 150 kVp. Other scan parameters were kept identical. The differences (∆) in mean HU and standard deviation (SD) in images, with and without metal, were measured and compared. In the clinical study, three independent readers performed visual image quality assessments on eight different protocols using retrospectively acquired images.

RESULTS

Iterative MAR had the lowest ∆HU and ∆SD in the phantom study. For images of the forearm, the soft tissue noise for Sn-based 150-kVp UHR protocol with was significantly higher (p = 0.037) than for single-energy MAR protocols. All Sn-based 150-kVp protocols were rated significantly higher (p < 0.046 than the single-energy MAR protocols in the visual assessment.

CONCLUSIONS

All Sn-based 150-kVp UHR protocols showed similar objective MAR in the phantom study, and higher objective MAR and significantly improved visual image quality than single-energy MAR.

RELEVANCE STATEMENT

Images with less metal artifacts and higher visual image quality may be more clinically optimal in CT examination of musculoskeletal patients with metal implants.

KEY POINTS

• Metal artifact reduction algorithms and Sn filter combined with high kVp reduce artifacts. • Metal artifact reduction algorithms introduce new artifacts in certain metals. • Sn-based protocols alone may be considered as low metal artifact protocols.

Visual and quantitative assessment of hip implant-related metal artifacts at low field MRI: a phantom study comparing a 0.55-T system with 1.5-T and 3-T systems

BACKGROUND

To investigate hip implant-related metal artifacts on a 0.55-T system compared with 1.5-T and 3-T systems.

METHODS

Total hip arthroplasty made of three different alloys were evaluated in a water phantom at 0.55, 1.5, and 3 T using routine protocols. Visually assessment (VA) was performed by three readers using a Likert scale from 0 (no artifacts) to 6 (extremely severe artifacts). Quantitative assessment (QA) was performed using the coefficient of variation (CoV) and the fraction of voxels within a threshold of the mean signal intensity compared to an automatically defined region of interest (FVwT). Agreement was evaluated using intra/inter-class correlation coefficient (ICC).

RESULTS

Interreader agreement of VA was strong-to-moderate (ICC 0.74-0.82). At all field strengths (0.55-T/1.5-T/3-T), artifacts were assigned a lower score for titanium (Ti) alloys (2.44/2.9/2.7) than for stainless steel (Fe-Cr) (4.1/3.9/5.1) and cobalt-chromium (Co-Cr) alloys (4.1/4.1/5.2) (p < 0.001 for both). Artifacts were lower for 0.55-T and 1.5-T than for 3-T systems, for all implants (p ≤ 0.049). A strong VA-to-QA correlation was found (r = 0.81; p < 0.001); CoV was lower for Ti alloys than for Fe-Cr and Co-Cr alloys at all field strengths. The FVwT showed a negative correlation with VA (-0.68 < r < -0.84; p < 0.001).

CONCLUSIONS

Artifact intensity was lowest for Ti alloys at 0.55 T. For other alloys, it was similar at 0.55 T and 1.5 T, higher at 3 T. Despite an inferior gradient system and a larger bore width, the 0.55-T system showed the same artifact intensity of the 1.5-T system.

Iron deposition in multiple sclerosis: overall load or distribution alteration?

Background

Though abnormal iron deposition has been reported in specific brain regions in multiple sclerosis (MS), no data exist about whether the overall quantity of iron in the brain is altered or not. We aimed to determine whether the noted aberrant iron deposition in MS brains was a problem of overall load or regional distribution in a cohort of MS patients.

Methods

An experienced neuroradiologist, a radiology software engineer, and four neurologists analysed data from quantitative susceptibility maps reconstructed from 3-T magnetic resonance brain images of 30 MS patients and 15 age- and sex-matched healthy controls. Global brain iron load was calculated, and the regional iron concentrations were assessed in 1,000 regions of interest placed in MS lesions in different locations, normal appearing white matter, thalami, and basal ganglia.

Results

Global brain iron load was comparable between patients and controls after adjustment for volume (p = 0.660), whereas the regional iron concentrations were significantly different in patients than in control (p ≤ 0.031). There was no significant correlation between global iron load and clinical parameters, whereas regional iron concentrations correlated with patients’ age, disease duration, and disability grade (p ≤ 0.039).

Conclusions

The aberrant iron deposition noted in MS seems to be a problem of regional distribution rather than an altered global brain iron load.

Supplementary Information

The online version contains supplementary material available at 10.1186/s41747-022-00279-9.

Artificial intelligence for prostate MRI: open datasets, available applications, and grand challenges

Artificial intelligence (AI) for prostate magnetic resonance imaging (MRI) is starting to play a clinical role for prostate cancer (PCa) patients. AI-assisted reading is feasible, allowing workflow reduction. A total of 3,369 multi-vendor prostate MRI cases are available in open datasets, acquired from 2003 to 2021 in Europe or USA at 3 T (n = 3,018; 89.6%) or 1.5 T (n = 296; 8.8%), 346 cases scanned with endorectal coil (10.3%), 3,023 (89.7%) with phased-array surface coils; 412 collected for anatomical segmentation tasks, 3,096 for PCa detection/classification; for 2,240 cases lesions delineation is available and 56 cases have matching histopathologic images; for 2,620 cases the PSA level is provided; the total size of all open datasets amounts to approximately 253 GB. Of note, quality of annotations provided per dataset highly differ and attention must be paid when using these datasets (e.g., data overlap). Seven grand challenges and commercial applications from eleven vendors are here considered. Few small studies provided prospective validation. More work is needed, in particular validation on large-scale multi-institutional, well-curated public datasets to test general applicability. Moreover, AI needs to be explored for clinical stages other than detection/characterization (e.g., follow-up, prognosis, interventions, and focal treatment).

Radiomics in photon-counting dedicated breast CT: potential of texture analysis for breast density classification

Background

We investigated whether features derived from texture analysis (TA) can distinguish breast density (BD) in spiral photon-counting breast computed tomography (PC-BCT).

Methods

In this retrospective single-centre study, we analysed 10,000 images from 400 PC-BCT examinations of 200 patients. Images were categorised into four-level density scale (a–d) using Breast Imaging Reporting and Data System (BI-RADS)-like criteria. After manual definition of representative regions of interest, 19 texture features (TFs) were calculated to analyse the voxel grey-level distribution in the included image area. ANOVA, cluster analysis, and multinomial logistic regression statistics were used. A human readout then was performed on a subset of 60 images to evaluate the reliability of the proposed feature set.

Results

Of the 19 TFs, 4 first-order features and 7 second-order features showed significant correlation with BD and were selected for further analysis. Multinomial logistic regression revealed an overall accuracy of 80% for BD assessment. The majority of TFs systematically increased or decreased with BD. Skewness (rho -0.81), as a first-order feature, and grey-level nonuniformity (GLN, -0.59), as a second-order feature, showed the strongest correlation with BD, independently of other TFs. Mean skewness and GLN decreased linearly from density a to d. Run-length nonuniformity (RLN), as a second-order feature, showed moderate correlation with BD, but resulted in redundant being correlated with GLN. All other TFs showed only weak correlation with BD (range -0.49 to 0.49, p < 0.001) and were neglected.

Conclusion

TA of PC-BCT images might be a useful approach to assess BD and may serve as an observer-independent tool.

Reproducibility of radiomic features in CT images of NSCLC patients: an integrative analysis on the impact of acquisition and reconstruction parameters

BACKGROUND

We investigated to what extent tube voltage, scanner model, and reconstruction algorithm affect radiomic feature reproducibility in a single-institution retrospective database of computed tomography images of non-small-cell lung cancer patients.

METHODS

This study was approved by the Institutional Review Board (UID 2412). Images of 103 patients were considered, being acquired on either among two scanners, at 100 or 120 kVp. For each patient, images were reconstructed with six iterative blending levels, and 1414 features were extracted from each reconstruction. At univariate analysis, Wilcoxon-Mann-Whitney test was applied to evaluate feature differences within scanners and voltages, whereas the impact of the reconstruction was established with the overall concordance correlation coefficient (OCCC). A multivariable mixed model was also applied to investigate the independent contribution of each acquisition/reconstruction parameter. Univariate and multivariable analyses were combined to analyse feature behaviour.

RESULTS

Scanner model and voltage did not affect features significantly. The reconstruction blending level showed a significant impact at both univariate analysis (154/1414 features yielding an OCCC < 0.85) and multivariable analysis, with most features (1042/1414) revealing a systematic trend with the blending level (multiple comparisons adjusted p < 0.05). Reproducibility increased in association to image processing with smooth filters, nonetheless specific investigation in relation to clinical endpoints should be performed to ensure that textural information is not removed.

CONCLUSIONS

Combining univariate and multivariable models is allowed to identify features for which corrections may be applied to reduce the trend with the algorithm and increase reproducibility. Subsequent clustering may be applied to eliminate residual redundancy.

Creating a training set for artificial intelligence from initial segmentations of airways

Airways segmentation is important for research about pulmonary disease but require a large amount of time by trained specialists. We used an openly available software to improve airways segmentations obtained from an artificial intelligence (AI) tool and retrained the tool to get a better performance. Fifteen initial airway segmentations from low-dose chest computed tomography scans were obtained with a 3D-Unet AI tool previously trained on Danish Lung Cancer Screening Trial and Erasmus-MC Sophia datasets. Segmentations were manually corrected in 3D Slicer. The corrected airway segmentations were used to retrain the 3D-Unet. Airway measurements were automatically obtained and included count, airway length and luminal diameter per generation from the segmentations. Correcting segmentations required 2–4 h per scan. Manually corrected segmentations had more branches (p < 0.001), longer airways (p < 0.001) and smaller luminal diameters (p = 0.004) than initial segmentations. Segmentations from retrained 3D-Unets trended towards more branches and longer airways compared to the initial segmentations. The largest changes were seen in airways from 6th generation onwards. Manual correction results in significantly improved segmentations and is potentially a useful and time-efficient method to improve the AI tool performance on a specific hospital or research dataset.

An evaluation of performance measures for arterial brain vessel segmentation

BACKGROUND

Arterial brain vessel segmentation allows utilising clinically relevant information contained within the cerebral vascular tree. Currently, however, no standardised performance measure is available to evaluate the quality of cerebral vessel segmentations. Thus, we developed a performance measure selection framework based on manual visual scoring of simulated segmentation variations to find the most suitable measure for cerebral vessel segmentation.

METHODS

To simulate segmentation variations, we manually created non-overlapping segmentation errors common in magnetic resonance angiography cerebral vessel segmentation. In 10 patients, we generated a set of approximately 300 simulated segmentation variations for each ground truth image. Each segmentation was visually scored based on a predefined scoring system and segmentations were ranked based on 22 performance measures common in the literature. The correlation of visual scores with performance measure rankings was calculated using the Spearman correlation coefficient.

RESULTS

The distance-based performance measures balanced average Hausdorff distance (rank = 1) and average Hausdorff distance (rank = 2) provided the segmentation rankings with the highest average correlation with manual rankings. They were followed by overlap-based measures such as Dice coefficient (rank = 7), a standard performance measure in medical image segmentation.

CONCLUSIONS

Average Hausdorff distance-based measures should be used as a standard performance measure in evaluating cerebral vessel segmentation quality. They can identify more relevant segmentation errors, especially in high-quality segmentations. Our findings have the potential to accelerate the validation and development of novel vessel segmentation approaches.

Tomographic similarity scan with a computed modified absolute mandibular midsagittal plane for precise and objective localization of mandibular asymmetry

The application of 3D imaging is at its cusp in craniofacial diagnosis and treatment planning. However, most applications are limited to simple subjective superimposition-based analysis. As the diagnostic accuracy dictates the precision in operability, we propose a novel method that enables objective clinical decision making for patients with mandibular asymmetry. We analyzed cone-beam computed tomography (CBCT) scans of 34 patients who underwent surgical correction for mandibular asymmetry using a high-throughput computing algorithm. Radiomic segmentation of quantitative features of surface and volume followed by exploration resulted in identification of a computed modified absolute mandibular midsagittal plane (cmAMP). Tomographic similarity scan (ToSS) curves were generated via bilateral equidistant scanning in an antero-posterior direction with cmAMP as the reference. ToSS comprised of a comprehensive similarity index (SI) score curve and a segment-wise volume curve. The SI score was computed using the Sørensen-Dice similarity coefficient ranging from 0 to 1. The volumetric analysis was represented as the non-overlapping volume (NOV) and overlapping volume (OV) for each segment, with two segmentation lines, at the mental foramen anteriorly and the intraoral vertical ramus osteotomy region posteriorly. Statistical analysis showed strong negative correlation between the NOV and SI scores for the anterior, middle, and total mandible (P < 0.001). Additionally, a significant correlation was observed between the change in the SI scores for anterior (P = 0.044) and middle segments (P < 0.001) to the total mandible when comparing the data before and after the surgery. This work demonstrated the potential of incorporating ToSS curves in surgical simulation software to improve precision in the clinical decision-making process.

On the usage of average Hausdorff distance for segmentation performance assessment: hidden error when used for ranking

Average Hausdorff distance is a widely used performance measure to calculate the distance between two point sets. In medical image segmentation, it is used to compare ground truth images with segmentations allowing their ranking. We identified, however, ranking errors of average Hausdorff distance making it less suitable for applications in segmentation performance assessment. To mitigate this error, we present a modified calculation of this performance measure that we have coined “balanced average Hausdorff distance”. To simulate segmentations for ranking, we manually created non-overlapping segmentation errors common in magnetic resonance angiography cerebral vessel segmentation as our use-case. Adding the created errors consecutively and randomly to the ground truth, we created sets of simulated segmentations with increasing number of errors. Each set of simulated segmentations was ranked using both performance measures. We calculated the Kendall rank correlation coefficient between the segmentation ranking and the number of errors in each simulated segmentation. The rankings produced by balanced average Hausdorff distance had a significantly higher median correlation (1.00) than those by average Hausdorff distance (0.89). In 200 total rankings, the former misranked 52 whilst the latter misranked 179 segmentations. Balanced average Hausdorff distance is more suitable for rankings and quality assessment of segmentations than average Hausdorff distance.

Impact of inter-reader contouring variability on textural radiomics of colorectal liver metastases

BACKGROUND

Radiomics is expected to improve the management of metastatic colorectal cancer (CRC). We aimed at evaluating the impact of liver lesion contouring as a source of variability on radiomic features (RFs).

METHODS

After Ethics Committee approval, 70 liver metastases in 17 CRC patients were segmented on contrast-enhanced computed tomography scans by two residents and checked by experienced radiologists. RFs from grey level co-occurrence and run length matrices were extracted from three-dimensional (3D) regions of interest (ROIs) and the largest two-dimensional (2D) ROIs. Inter-reader variability was evaluated with Dice coefficient and Hausdorff distance, whilst its impact on RFs was assessed using mean relative change (MRC) and intraclass correlation coefficient (ICC). For the main lesion of each patient, one reader also segmented a circular ROI on the same image used for the 2D ROI.

RESULTS

The best inter-reader contouring agreement was observed for 2D ROIs according to both Dice coefficient (median 0.85, interquartile range 0.78-0.89) and Hausdorff distance (0.21 mm, 0.14-0.31 mm). Comparing RF values, MRC ranged 0-752% for 2D and 0-1567% for 3D. For 24/32 RFs (75%), MRC was lower for 2D than for 3D. An ICC > 0.90 was observed for more RFs for 2D (53%) than for 3D (34%). Only 2/32 RFs (6%) showed a variability between 2D and circular ROIs higher than inter-reader variability.

CONCLUSIONS

A 2D contouring approach may help mitigate overall inter-reader variability, albeit stable RFs can be extracted from both 3D and 2D segmentations of CRC liver metastases.

Artificial intelligence in bone age assessment: accuracy and efficiency of a novel fully automated algorithm compared to the Greulich-Pyle method

Background

Bone age (BA) assessment performed by artificial intelligence (AI) is of growing interest due to improved accuracy, precision and time efficiency in daily routine. The aim of this study was to investigate the accuracy and efficiency of a novel AI software version for automated BA assessment in comparison to the Greulich-Pyle method.

Methods

Radiographs of 514 patients were analysed in this retrospective study. Total BA was assessed independently by three blinded radiologists applying the GP method and by the AI software. Overall and gender-specific BA assessment results, as well as reading times of both approaches, were compared, while the reference BA was defined by two blinded experienced paediatric radiologists in consensus by application of the Greulich-Pyle method.

Results

Mean absolute deviation (MAD) and root mean square deviation (RSMD) were significantly lower between AI-derived BA and reference BA (MAD 0.34 years, RSMD 0.38 years) than between reader-calculated BA and reference BA (MAD 0.79 years, RSMD 0.89 years; p < 0.001). The correlation between AI-derived BA and reference BA (r = 0.99) was significantly higher than between reader-calculated BA and reference BA (r = 0.90; p < 0.001). No statistical difference was found in reader agreement and correlation analyses regarding gender (p = 0.241). Mean reading times were reduced by 87% using the AI system.

Conclusions

A novel AI software enabled highly accurate automated BA assessment. It may improve efficiency in clinical routine by reducing reading times without compromising the accuracy compared with the Greulich-Pyle method.

Radiomics: the facts and the challenges of image analysis

Radiomics is an emerging translational field of research aiming to extract mineable high-dimensional data from clinical images. The radiomic process can be divided into distinct steps with definable inputs and outputs, such as image acquisition and reconstruction, image segmentation, features extraction and qualification, analysis, and model building. Each step needs careful evaluation for the construction of robust and reliable models to be transferred into clinical practice for the purposes of prognosis, non-invasive disease tracking, and evaluation of disease response to treatment. After the definition of texture parameters (shape features; first-, second-, and higher-order features), we briefly discuss the origin of the term radiomics and the methods for selecting the parameters useful for a radiomic approach, including cluster analysis, principal component analysis, random forest, neural network, linear/logistic regression, and other. Reproducibility and clinical value of parameters should be firstly tested with internal cross-validation and then validated on independent external cohorts. This article summarises the major issues regarding this multi-step process, focussing in particular on challenges of the extraction of radiomic features from data sets provided by computed tomography, positron emission tomography, and magnetic resonance imaging.

Comparison of DTI analysis methods for clinical research: influence of pre-processing and tract selection methods

Background

The primary aim was to compare fractional anisotropy (FA) values derived with different diffusion tensor imaging (DTI) analysis approaches (atlas-based, streamline tractography, and combined). A secondary aim was to compare FA values and number of tracts (NT) with the clinical motor outcome quantified by the functional independence measure for children (WeeFIM).

Methods

Thirty-nine DTI datasets of children with acquired brain injury were analysed. Regions of interest for the ipsilesional corticospinal tract were defined and mean FA and NT were calculated. We evaluated FA values with Spearman correlation, the Friedman and Wilcoxon tests, and Bland-Altman analysis. DTI values were compared to WeeFIM values by non-parametric partial correlation and accuracy was assessed by receiver operating characteristics analysis.

Results

The FA values from all approaches correlated significantly with each other (p < 0.001). However, the FA values from streamline tractography were significantly higher (mean ± standard deviation (SD), 0.52 ± 0.08) than those from the atlas-based (0.42 ± 0.11) or the combined approach (0.41 ± 0.11) (p < 0.001 for both). FA and NT values correlated significantly with WeeFIM values (atlas-based FA, partial correlation coefficient (ρ) = 0.545, p = 0.001; streamline FA, ρ = 0.505, p = 0.002; NT, ρ = 0.434, p = 0.008; combined FA, ρ = 0.611, p < 0.001). FA of the atlas-based approach (sensitivity 90%, specificity 67%, area under the curve 0.82) and the combined approach (87%, 67%, 0.82), provided the highest predictive accuracy for outcome compared to FA (70%, 67%, 0.67) and NT (50%, 100%, 0.79, respectively) of the streamline approach.

Conclusion

FA values from streamline tractography were higher than those from the atlas-based and combined approach. The atlas-based and combined approach offer the best predictive accuracy for motor outcome, although both atlas-based and streamline tractography approaches provide significant predictors of clinical outcome.

Radiomics: the facts and the challenges of image analysis

Radiomics is an emerging translational field of research aiming to extract mineable high-dimensional data from clinical images. The radiomic process can be divided into distinct steps with definable inputs and outputs, such as image acquisition and reconstruction, image segmentation, features extraction and qualification, analysis, and model building. Each step needs careful evaluation for the construction of robust and reliable models to be transferred into clinical practice for the purposes of prognosis, non-invasive disease tracking, and evaluation of disease response to treatment. After the definition of texture parameters (shape features; first-, second-, and higher-order features), we briefly discuss the origin of the term radiomics and the methods for selecting the parameters useful for a radiomic approach, including cluster analysis, principal component analysis, random forest, neural network, linear/logistic regression, and other. Reproducibility and clinical value of parameters should be firstly tested with internal cross-validation and then validated on independent external cohorts. This article summarises the major issues regarding this multi-step process, focussing in particular on challenges of the extraction of radiomic features from data sets provided by computed tomography, positron emission tomography, and magnetic resonance imaging.