Current Applications and Future Impact of Machine Learning in Radiology

Choy, Garry; Khalilzadeh, Omid; Michalski, Mark; Do, Synho; Samir, Anthony E.; Pianykh, Oleg S.; Geis, J. Raymond; Pandharipande, Pari V.; Brink, James A.; Dreyer, Keith J.

doi:10.1148/radiol.2018171820

Cited by 638 publications

(526 citation statements)

References 88 publications

Supporting

Mentioning

493

Contrasting

Unclassified

Order By: Relevance

“…Alternatively, pedagogical support might be offered in specialized national centers, and some international training courses are starting to be offered to the bone research community. For the future, it may be expected that machine learning and computer‐aided diagnosis will be successfully applied to this problem …”

Section: Observer Variablesmentioning

confidence: 99%

“…For the future, it may be expected that machine learning and computer-aided diagnosis will be successfully applied to this problem. (48) Oei and colleagues (43) (a) Aim: estimate statistical measures of agreement and prevalence of osteoporotic vertebral fractures across the two most commonly applied assessment methods (b) OVF defined by a team of seven trained research assistants applied readings using a commercial morphometric tool (35) (QM), a second team of another seven trained research assistants, applied the algorithm designed described above (ABQ), with definite or uncertain findings reviewed by a musculoskeletal radiologist) (Morphologic) (a) n ¼ 7582 (both sexes) (b) Interrater agreement was moderate for both methods and intramethod agreement was poor (c) The prevalence of ABQ-defined OVF was lower compared to QM-defined OVF (d) ABQ OVF were more strongly associated with BMD and future fracture risk compared to QM (e) ABQ-defined OVF were mostly located at the thoracolumbar junction whereas QM OVF were mostly located at the midthoracic region (f) Intermethod agreement increased when excluding mild deformities for the definition of QM or when reassessing mild QM for endplate depression Mild deformities should be assessed for endplate depression, thereby decreasing false-positive QM fractures and reconciling the two methods Lentle and colleagues (44) (a) Aim: to compare the GSQ and an mABQ tool for radiologic (b) OVF defined by radiologists (including a musculoskeletal rad.) applying (1) GSQ method above and (2) same radiologists strictly applying an algorithm (modified after Jiang and colleagues (39) ; mABQ) designed to exclude false positives (a) n ¼ 6236 (both sexes) (b) Observer agreement was higher for mABQ compared to GSQ (c) The prevalence of OVF was lower when defined with mABQ compared to GSQ (d) mABQ-defined OVFs were more strongly associated with low BMD compared to GSQ (e) Prevalent mABQ were more strongly associated with incident OVF and non-OVF fractures Defining OVF by mABQ is preferred to the use of GSQ for clinical assessments Deng and colleagues (45) (a) Aim: to evaluate ECF-based method for detecting OVF in elderly Chinese population (a) n ¼ 3907 (both sexes) (b) ECF-defined OVFs are associated with lower BMD compared to GSQ-defined OVFs ECF might be more specific to assess mild OVF…”

Section: Observer Variablesmentioning

confidence: 99%

See 1 more Smart Citation

The Radiology of Osteoporotic Vertebral Fractures Revisited

Lentle¹,

Koromani

Brown³

et al. 2019

Journal of Bone and Mineral Research

View full text Add to dashboard Cite

Until recently there has been little evidence available to validate any method by which to make an accurate diagnosis of an osteoporotic vertebral fractures (OVFs) from plain radiographs. In part this reflects a lack of a completely satisfactory “gold standard,” but primarily it relates to the absence of well‐designed prospective studies in this context. Historically, OVFs were recognized by evidence of macroscopic structural failure in vertebrae using the criteria applied elsewhere in the skeleton. This comprised altered alignment, fragmentation, cortical disruptions, and breaks, among other changes. However, these morphological criteria were replaced by vertebral morphometry, referring to the use of quantitative or quasi‐quantitative measurement tools for fracture diagnosis. Vertebral morphometry emerged as an understanding of and treatment for osteoporosis evolved, mainly in response to the need for expeditious assessments of large numbers of spine images for epidemiological and pharmaceutical purposes. Although most of the descriptions of such morphometric tools have stressed that they were not to be applied to clinical diagnosis with respect to individual patients, this constraint has been widely disregarded. Here we review the major attempts to develop a diagnostic strategy for OVF and describe their characteristics in adults and children. Recent evidence suggests that morphometric (quantitative; ie, based on measurement of dimensions and shape description) criteria are inferior to morphologic (qualitative; ie, based on structural integrity) vertebral damage assessment in identifying people with low bone density and at an increased risk of future fracture. Thus there is now an evidentiary basis for suggesting that morphological assessment is the preferred strategy for use in diagnosing OVF from radiographs. © 2019 American Society for Bone and Mineral Research.

show abstract

Section: Observer Variablesmentioning

confidence: 99%

Section: Observer Variablesmentioning

confidence: 99%

The Radiology of Osteoporotic Vertebral Fractures Revisited

Lentle¹,

Koromani

Brown³

et al. 2019

Journal of Bone and Mineral Research

View full text Add to dashboard Cite

show abstract

“…31,32 This approach may result in increasing workflow efficiency of an echocardiography laboratory. 33,34 Particularly, high-volume centers may take advantage of this new development to save time, while low-volume centers may reach higher standards in terms of quality of their measurements.…”

Section: Discussionmentioning

confidence: 99%

Automated, machine learning‐based, 3D echocardiographic quantification of left ventricular mass

et al. 2018

View full text Add to dashboard Cite

Background: Although 3D echocardiography (3DE) circumvents many limitations of 2D echocardiography by allowing direct measurements of left ventricular (LV) mass, it is seldom used in clinical practice due to time-consuming analysis. A recently developed 3DE machine learning (ML) approach allows automated determination of LV mass. We aimed to evaluate the accuracy of this new approach by comparing it to cardiac magnetic resonance (CMR) reference and to conventional 3DE volumetric analysis. Methods:We prospectively studied 23 patients who underwent 3DE (Philips EPIQ) and CMR imaging on the same day. Single-beat wide-angle 3D datasets of the left ventricle were acquired. LV mass was quantified using the new automated software (Philips HeartModel) with manual corrections when necessary and using conventional volumetric analysis (TomTec). CMR analysis was performed by manual slice-byslice tracing of LV endo-and epicardial boundaries. Reproducibility of the ML approach was assessed using repeated measurements and quantified by intra-class correlation (ICC) and coefficients of variation (CoV).Results: Automated LV mass measurements were feasible in 20 patients (87%). The results were similar to CMR-derived values (Bland-Altman bias 5 g, limits of agreement ±37 g) and also to the conventional 3DE analysis (bias 7 g, ±27 g). Processing time was considerably shorter: 1.02 ± 0.24 minutes (CMR: 2.20 ± 0.13 minutes;TomTec: 2.36 ± 0.09 minutes), although manual corrections were performed in most patients. Repeated measurements showed high reproducibility: ICC = 0.99; CoV = 4 ± 5%.Conclusions: 3D Echocardiography analysis of LV mass using novel ML-based algorithm is feasible, fast, and accurate and may thus facilitate the incorporation of 3DE measurements of LV mass into clinical practice. K E Y W O R D S2D echocardiography, 3D echocardiography, cardiac magnetic resonance imaging, left ventricular mass, machine learning | 313 VOLPATO eT AL.

show abstract

“…In recent years, deep learning utilizing convolutional neural networks, a form of machine learning, has been successfully employed to accurately perform a variety of image recognition and classification tasks by encoding hierarchies of spatial features through adaptive mathematical models [11][12][13][14]. Emerging medical applications range from automated bone age assessment to automated CXR diagnosis [11][12][13][14][15][16][17][18][19][20][21]. Deep learning thus also has potential to automate Brasfield scoring, thereby reducing the need for subspecialized readers to perform tedious processes while maintaining reliable quantitative metrics.…”

Section: Introductionmentioning

confidence: 99%

Deep learning to automate Brasfield chest radiographic scoring for cystic fibrosis

Barnes

Lungren

Shpanskaya

et al. 2020

Journal of Cystic Fibrosis

View full text Add to dashboard Cite

a b s t r a c tBackground: The aim of this study was to evaluate the hypothesis that a deep convolutional neural network (DCNN) model could facilitate automated Brasfield scoring of chest radiographs (CXRs) for patients with cystic fibrosis (CF), performing similarly to a pediatric radiologist. Methods: All frontal/lateral chest radiographs (2058 exams) performed in CF patients at a single institution from January 2008-2018 were retrospectively identified, and ground-truth Brasfield scoring performed by a board-certified pediatric radiologist. 1858 exams (90.3%) were used to train and validate the DCNN model, while 200 exams (9.7%) were reserved for a test set. Five board-certified pediatric radiologists independently scored the test set according to the Brasfield method. DCNN model vs. radiologist performance was compared using Spearman correlation ( ρ) as well as mean difference (MD), mean absolute difference (MAD), and root mean squared error (RMSE) estimation.Results: For the total Brasfield score, ρ for the model-derived results computed pairwise with each radiologist's scores ranged from 0.79-0.83, compared to 0.85-0.90 for radiologist vs. radiologist scores. The MD between model estimates of the total Brasfield score and the average score of radiologists was −0.09. Based on MD, MAD, and RMSE, the model matched or exceeded radiologist performance for all subfeatures except air-trapping and large lesions.Conclusions: A DCNN model is promising for predicting CF Brasfield scores with accuracy similar to that of a pediatric radiologist.

show abstract

Current Applications and Future Impact of Machine Learning in Radiology

Cited by 638 publications

References 88 publications

The Radiology of Osteoporotic Vertebral Fractures Revisited

The Radiology of Osteoporotic Vertebral Fractures Revisited

Automated, machine learning‐based, 3D echocardiographic quantification of left ventricular mass

Deep learning to automate Brasfield chest radiographic scoring for cystic fibrosis

Contact Info

Product

Resources

About