Strategies for Implementing Machine Learning Algorithms in the Clinical Practice of Radiology.

Despite recent advancements in machine learning (ML) applications in health care, there have been few benefits and improvements to clinical medicine in the hospital setting. To facilitate clinical adaptation of methods in ML, this review proposes a standardized framework for the step-by-step implementation of artificial intelligence into the clinical practice of radiology that focuses on three key components: problem identification, stakeholder alignment, and pipeline integration. A review of the recent literature and empirical evidence in radiologic imaging applications justifies this approach and offers a discussion on structuring implementation efforts to help other hospital practices leverage ML to improve patient care. Clinical trial registration no. 04242667 © RSNA, 2024 Supplemental material is available for this article.

Deep learning approach for automated segmentation of myocardium using bone scintigraphy single-photon emission computed tomography/computed tomography in patients with suspected cardiac amyloidosis.

BACKGROUND: We employed deep learning to automatically detect myocardial bone-seeking uptake as a marker of transthyretin cardiac amyloid cardiomyopathy (ATTR-CM) in patients undergoing 99mTc-pyrophosphate (PYP) or hydroxydiphosphonate (HDP) single-photon emission computed tomography (SPECT)/computed tomography (CT). METHODS: We identified a primary cohort of 77 subjects at Brigham and Women's Hospital and a validation cohort of 93 consecutive patients imaged at the University of Pennsylvania who underwent SPECT/CT with PYP and HDP, respectively, for evaluation of ATTR-CM. Global heart regions of interest (ROIs) were traced on CT axial slices from the apex of the ventricle to the carina. Myocardial images were visually scored as grade 0 (no uptake), 1 (uptakeribs). A 2D U-net architecture was used to develop whole-heart segmentations for CT scans. Uptake was determined by calculating a heart-to-blood pool (HBP) ratio between the maximal counts value of the total heart region and the maximal counts value of the most superior ROI. RESULTS: Deep learning and ground truth segmentations were comparable (p=0.63). A total of 42 (55%) patients had abnormal myocardial uptake on visual assessment. Automated quantification of the mean HBP ratio in the primary cohort was 3.1±1.4 versus 1.4±0.2 (p<0.01) for patients with positive and negative cardiac uptake, respectively. The model had 100% accuracy in the primary cohort and 98% in the validation cohort. CONCLUSION: We have developed a highly accurate diagnostic tool for automatically segmenting and identifying myocardial uptake suggestive of ATTR-CM.

CT imaging-derived phenotypes for abdominal muscle and their association with age and sex in a medical biobank.

The study of muscle mass as an imaging-derived phenotype (IDP) may yield new insights into determining the normal and pathologic variations in muscle mass in the population. This can be done by determining 3D abdominal muscle mass from 12 distinct abdominal muscle regions and groups using computed tomography (CT) in a racially diverse medical biobank. To develop a fully automatic technique for assessment of CT abdominal muscle IDPs and preliminarily determine abdominal muscle IDP variations with age and sex in a clinically and racially diverse medical biobank. This retrospective study was conducted using the Penn Medicine BioBank (PMBB), a research protocol that recruits adult participants during outpatient visits at hospitals in the Penn Medicine network. We developed a deep residual U-Net (ResUNet) to segment 12 abdominal muscle groups including the left and right psoas, quadratus lumborum, erector spinae, gluteus medius, rectus abdominis, and lateral abdominals. 110 CT studies were randomly selected for training, validation, and testing. 44 of the 110 CT studies were selected to enrich the dataset with representative cases of intra-abdominal and abdominal wall pathology. The studies were divided into non-overlapping training, validation and testing sets. Model performance was evaluated using the Sørensen-Dice coefficient. Volumes of individual muscle groups were plotted to distribution curves. To investigate associations between muscle IDPs, age, and sex, deep learning model segmentations were performed on a larger abdominal CT dataset from PMBB consisting of 295 studies. Multivariable models were used to determine relationships between muscle mass, age and sex. The model's performance (Dice scores) on the test data was the following: psoas: 0.85 ± 0.12, quadratus lumborum: 0.72 ± 0.14, erector spinae: 0.92 ± 0.07, gluteus medius: 0.90 ± 0.08, rectus abdominis: 0.85 ± 0.08, lateral abdominals: 0.85 ± 0.09. The average Dice score across all muscle groups was 0.86 ± 0.11. Average total muscle mass for females was 2041 ± 560.7 g with a high of 2256 ± 560.1 g (41-50 year old cohort) and a change of - 0.96 g/year, declining to an average mass of 1579 ± 408.8 g (81-100 year old cohort). Average total muscle mass for males was 3086 ± 769.1 g with a high of 3385 ± 819.3 g (51-60 year old cohort) and a change of - 1.73 g/year, declining to an average mass of 2629 ± 536.7 g (81-100 year old cohort). Quadratus lumborum was most highly correlated with age for both sexes (correlation coefficient of - 0.5). Gluteus medius mass in females was positively correlated with age with a coefficient of 0.22. These preliminary findings show that our CNN can automate detailed abdominal muscle volume measurement. Unlike prior efforts, this technique provides 3D muscle segmentations of individual muscles. This technique will dramatically impact sarcopenia diagnosis and research, elucidating its clinical and public health implications. Our results suggest a peak age range for muscle mass and an expected rate of decline, both of which vary between genders. Future goals are to investigate genetic variants for sarcopenia and malnutrition, while describing genotype-phenotype associations of muscle mass in healthy humans using imaging-derived phenotypes. It is feasible to obtain 3D abdominal muscle IDPs with high accuracy from patients in a medical biobank using fully automated machine learning methods. Abdominal muscle IDPs showed significant variations in lean mass by age and sex. In the future, this tool can be leveraged to perform a genome-wide association study across the medical biobank and determine genetic variants associated with early or accelerated muscle wasting.

Clinical correlates of CT imaging-derived phenotypes among lean and overweight patients with hepatic steatosis.

The objective of this study is to define CT imaging derived phenotypes for patients with hepatic steatosis, a common metabolic liver condition, and determine its association with patient data from a medical biobank. There is a need to further characterize hepatic steatosis in lean patients, as its epidemiology may differ from that in overweight patients. A deep learning method determined the spleen-hepatic attenuation difference (SHAD) in Hounsfield Units (HU) on abdominal CT scans as a quantitative measure of hepatic steatosis. The patient cohort was stratified by BMI with a threshold of 25 kg/m2 and hepatic steatosis with threshold SHAD ≥ - 1 HU or liver mean attenuation ≤ 40 HU. Patient characteristics, diagnoses, and laboratory results representing metabolism and liver function were investigated. A phenome-wide association study (PheWAS) was performed for the statistical interaction between SHAD and the binary characteristic LEAN. The cohort contained 8914 patients-lean patients with (N = 278, 3.1%) and without (N = 1867, 20.9%) steatosis, and overweight patients with (N = 1863, 20.9%) and without (N = 4906, 55.0%) steatosis. Among all lean patients, those with steatosis had increased rates of cardiovascular disease (41.7 vs 27.8%), hypertension (86.7 vs 49.8%), and type 2 diabetes mellitus (29.1 vs 15.7%) (all p < 0.0001). Ten phenotypes were significant in the PheWAS, including chronic kidney disease, renal failure, and cardiovascular disease. Hepatic steatosis was found to be associated with cardiovascular, kidney, and metabolic conditions, separate from overweight BMI.

SynthA1c: Towards Clinically Interpretable Patient Representations for Diabetes Risk Stratification.

Early diagnosis of Type 2 Diabetes Mellitus (T2DM) is crucial to enable timely therapeutic interventions and lifestyle modifications. As the time available for clinical office visits shortens and medical imaging data become more widely available, patient image data could be used to opportunistically identify patients for additional T2DM diagnostic workup by physicians. We investigated whether image-derived phenotypic data could be leveraged in tabular learning classifier models to predict T2DM risk in an automated fashion to flag high-risk patients without the need for additional blood laboratory measurements. In contrast to traditional binary classifiers, we leverage neural networks and decision tree models to represent patient data as 'SynthA1c' latent variables, which mimic blood hemoglobin A1c empirical lab measurements, that achieve sensitivities as high as 87.6%. To evaluate how SynthA1c models may generalize to other patient populations, we introduce a novel generalizable metric that uses vanilla data augmentation techniques to predict model performance on input out-of-domain covariates. We show that image-derived phenotypes and physical examination data together can accurately predict diabetes risk as a means of opportunistic risk stratification enabled by artificial intelligence and medical imaging. Our code is available at https://github.com/allisonjchae/DMT2RiskAssessment.

Automated, calibration-free quantification of cortical bone porosity and geometry in postmenopausal osteoporosis from ultrashort echo time MRI and deep learning.

BACKGROUND: Assessment of cortical bone porosity and geometry by imaging in vivo can provide useful information about bone quality that is independent of bone mineral density (BMD). Ultrashort echo time (UTE) MRI techniques of measuring cortical bone porosity and geometry have been extensively validated in preclinical studies and have recently been shown to detect impaired bone quality in vivo in patients with osteoporosis. However, these techniques rely on laborious image segmentation, which is clinically impractical. Additionally, UTE MRI porosity techniques typically require long scan times or external calibration samples and elaborate physics processing, which limit their translatability. To this end, the UTE MRI-derived Suppression Ratio has been proposed as a simple-to-calculate, reference-free biomarker of porosity which can be acquired in clinically feasible acquisition times. PURPOSE: To explore whether a deep learning method can automate cortical bone segmentation and the corresponding analysis of cortical bone imaging biomarkers, and to investigate the Suppression Ratio as a fast, simple, and reference-free biomarker of cortical bone porosity. METHODS: In this retrospective study, a deep learning 2D U-Net was trained to segment the tibial cortex from 48 individual image sets comprised of 46 slices each, corresponding to 2208 training slices. Network performance was validated through an external test dataset comprised of 28 scans from 3 groups: (1) 10 healthy, young participants, (2) 9 postmenopausal, non-osteoporotic women, and (3) 9 postmenopausal, osteoporotic women. The accuracy of automated porosity and geometry quantifications were assessed with the coefficient of determination and the intraclass correlation coefficient (ICC). Furthermore, automated MRI biomarkers were compared between groups and to dual energy X-ray absorptiometry (DXA)- and peripheral quantitative CT (pQCT)-derived BMD. Additionally, the Suppression Ratio was compared to UTE porosity techniques based on calibration samples. RESULTS: The deep learning model provided accurate labeling (Dice score 0.93, intersection-over-union 0.88) and similar results to manual segmentation in quantifying cortical porosity (R2 ≥ 0.97, ICC ≥ 0.98) and geometry (R2 ≥ 0.82, ICC ≥ 0.75) parameters in vivo. Furthermore, the Suppression Ratio was validated compared to established porosity protocols (R2 ≥ 0.78). Automated parameters detected age- and osteoporosis-related impairments in cortical bone porosity (P ≤ .002) and geometry (P values ranging from <0.001 to 0.08). Finally, automated porosity markers showed strong, inverse Pearson's correlations with BMD measured by pQCT (|R| ≥ 0.88) and DXA (|R| ≥ 0.76) in postmenopausal women, confirming that lower mineral density corresponds to greater porosity. CONCLUSION: This study demonstrated feasibility of a simple, automated, and ionizing-radiation-free protocol for quantifying cortical bone porosity and geometry in vivo from UTE MRI and deep learning.

Exome-wide association analysis of CT imaging-derived hepatic fat in a medical biobank.

Nonalcoholic fatty liver disease is common and highly heritable. Genetic studies of hepatic fat have not sufficiently addressed non-European and rare variants. In a medical biobank, we quantitate hepatic fat from clinical computed tomography (CT) scans via deep learning in 10,283 participants with whole-exome sequences available. We conduct exome-wide associations of single variants and rare predicted loss-of-function (pLOF) variants with CT-based hepatic fat and perform cross-modality replication in the UK Biobank (UKB) by linking whole-exome sequences to MRI-based hepatic fat. We confirm single variants previously associated with hepatic fat and identify several additional variants, including two (FGD5 H600Y and CITED2 S198_G199del) that replicated in UKB. A burden of rare pLOF variants in LMF2 is associated with increased hepatic fat and replicates in UKB. Quantitative phenotypes generated from clinical imaging studies and intersected with genomic data in medical biobanks have the potential to identify molecular pathways associated with human traits and disease.

A multiancestry genome-wide association study of unexplained chronic ALT elevation as a proxy for nonalcoholic fatty liver disease with histological and radiological validation.

Nonalcoholic fatty liver disease (NAFLD) is a growing cause of chronic liver disease. Using a proxy NAFLD definition of chronic elevation of alanine aminotransferase (cALT) levels without other liver diseases, we performed a multiancestry genome-wide association study (GWAS) in the Million Veteran Program (MVP) including 90,408 cALT cases and 128,187 controls. Seventy-seven loci exceeded genome-wide significance, including 25 without prior NAFLD or alanine aminotransferase associations, with one additional locus identified in European American-only and two in African American-only analyses (P < 5 × 10-8). External replication in histology-defined NAFLD cohorts (7,397 cases and 56,785 controls) or radiologic imaging cohorts (n = 44,289) replicated 17 single-nucleotide polymorphisms (SNPs) (P < 6.5 × 10-4), of which 9 were new (TRIB1, PPARG, MTTP, SERPINA1, FTO, IL1RN, COBLL1, APOH and IFI30). Pleiotropy analysis showed that 61 of 77 multiancestry and all 17 replicated SNPs were jointly associated with metabolic and/or inflammatory traits, revealing a complex model of genetic architecture. Our approach integrating cALT, histology and imaging reveals new insights into genetic liability to NAFLD.

Optimizing Medical Student Clerkship Schedules Using a Novel Application of the Hungarian Algorithm.

PROBLEM: Medical students often have preferences regarding the order of their clinical rotations, but assigning rotations fairly and efficiently can be challenging. To achieve a solution that optimizes assignments (i.e., maximizes student satisfaction), the authors present a novel application of the Hungarian algorithm, designed at the University of Texas Southwestern Medical Center (UTSW), to assign student schedules. APPROACH: Possible schedules were divided into distinct pathway options with k total number of seats. Each of n students submitted a ranked list of their top 5 pathway choices. An n × k matrix was formed, where the location (i, j) represented the cost associated with student i being placed in seat j. Progressively higher costs were assigned to students receiving less desired pathways. The Hungarian algorithm was then used to find the assignments that minimize total cost. The authors compared the performance of the Hungarian algorithm against 2 alternative algorithms (i.e., the rank and lottery algorithms). To evaluate the 3 algorithms, 4 simulations were conducted with different popularity weights for different pathways and were run across 1,000 trials. The algorithms were also compared using 3 years of UTSW student preference data for the classes of 2019, 2020, and 2021. OUTCOMES: In all 4 computer simulations, the Hungarian algorithm resulted in more students receiving 1 of their top 3 choices and fewer students receiving none of their preferences. Similarly, for UTSW student preference data, the Hungarian algorithm resulted in more students receiving 1 of their top 3 preferences and fewer students receiving none of their ranked preferences. NEXT STEPS: This approach may be broadly applied to scheduling challenges in undergraduate and graduate medical education. Furthermore, by manipulating cost values, additional constraints can be enforced (e.g., requiring certain seats to be filled, attempting to avoid schedules that begin with a student's desired specialty).

Classification of Myocardial 18F-FDG PET Uptake Patterns Using Deep Learning.

PURPOSE: To perform automated myocardial segmentation and uptake classification from whole-body fluorine 18 fluorodeoxyglucose (FDG) PET. MATERIALS AND METHODS: In this retrospective study, consecutive patients who underwent FDG PET imaging for oncologic indications were included (July-August 2018). The left ventricle (LV) on whole-body FDG PET images was manually segmented and classified as showing no myocardial uptake, diffuse uptake, or partial uptake. A total of 609 patients (mean age, 64 years ± 14 [standard deviation]; 309 women) were included and split between training (60%, 365 patients), validation (20%, 122 patients), and testing (20%, 122 patients) datasets. Two sequential neural networks were developed to automatically segment the LV and classify the myocardial uptake pattern using segmentation and classification training data provided by human experts. Linear regression was performed to correlate findings from human experts and deep learning. Classification performance was evaluated using receiver operating characteristic (ROC) analysis. RESULTS: There was moderate agreement of uptake pattern between experts and deep learning (as a fraction of correctly categorized images) with 78% (36 of 46) for no uptake, 71% (34 of 48) for diffuse uptake, and 71% (20 of 28) for partial uptake. There was no bias in LV volume for partial or diffuse uptake categories (P = .56); however, deep learning underestimated LV volumes in the no uptake category. There was good correlation for LV volume (R 2 = 0.35, b = .71). ROC analysis showed the area under the curve for classifying no uptake and diffuse uptake was high (> 0.90) but lower for partial uptake (0.77). The feasibility of a myocardial uptake index (MUI) for quantifying the degree of myocardial activity patterns was shown, and there was excellent visual agreement between MUI and uptake patterns. CONCLUSION: Deep learning was able to segment and classify myocardial uptake patterns on FDG PET images.Keywords: PET, Heart, Computer Aided Diagnosis, Computer Application-Detection/DiagnosisSupplemental material is available for this article.©RSNA, 2021.

Quantification of abdominal fat from computed tomography using deep learning and its association with electronic health records in an academic biobank.

OBJECTIVE: The objective was to develop a fully automated algorithm for abdominal fat segmentation and to deploy this method at scale in an academic biobank. MATERIALS AND METHODS: We built a fully automated image curation and labeling technique using deep learning and distributive computing to identify subcutaneous and visceral abdominal fat compartments from 52,844 computed tomography scans in 13,502 patients in the Penn Medicine Biobank (PMBB). A classification network identified the inferior and superior borders of the abdomen, and a segmentation network differentiated visceral and subcutaneous fat. Following technical evaluation of our method, we conducted studies to validate known relationships with visceral and subcutaneous fat. RESULTS: When compared with 100 manually annotated cases, the classification network was on average within one 5-mm slice for both the superior (0.4 ± 1.1 slice) and inferior (0.4 ± 0.6 slice) borders. The segmentation network also demonstrated excellent performance with intraclass correlation coefficients of 1.00 (P < 2 × 10-16) for subcutaneous and 1.00 (P < 2 × 10-16) for visceral fat on 100 testing cases. We performed integrative analyses of abdominal fat with the phenome extracted from the electronic health record and found highly significant associations with diabetes mellitus, hypertension, and renal failure, among other phenotypes. CONCLUSIONS: This work presents a fully automated and highly accurate method for the quantification of abdominal fat that can be applied to routine clinical imaging studies to fuel translational scientific discovery.