Deep learning workflow in radiology: a primer Montagnon, Emmanuel; Cerny, Milena; Cadrin-Chênevert, Alexandre ...
Insights into imaging,
02/2020, Letnik:
11, Številka:
1
Journal Article
Recenzirano
Odprti dostop
Interest for deep learning in radiology has increased tremendously in the past decade due to the high achievable performance for various computer vision tasks such as detection, segmentation, ...classification, monitoring, and prediction. This article provides step-by-step practical guidance for conducting a project that involves deep learning in radiology, from defining specifications, to deployment and scaling. Specifically, the objectives of this article are to provide an overview of clinical use cases of deep learning, describe the composition of multi-disciplinary team, and summarize current approaches to patient, data, model, and hardware selection. Key ideas will be illustrated by examples from a prototypical project on imaging of colorectal liver metastasis. This article illustrates the workflow for liver lesion detection, segmentation, classification, monitoring, and prediction of tumor recurrence and patient survival. Challenges are discussed, including ethical considerations, cohorting, data collection, anonymization, and availability of expert annotations. The practical guidance may be adapted to any project that requires automated medical image analysis.
Objective
To evaluate the diagnostic performance of intravoxel incoherent motion (IVIM) diffusion-weighted imaging (DWI) for grading hepatic inflammation.
Methods
In this retrospective ...cross-sectional dual-center study, 91 patients with chronic liver disease were recruited between September 2014 and September 2018. Patients underwent 3.0-T MRI examinations within 6 weeks from a liver biopsy. IVIM parameters, perfusion fraction (
f
), diffusion coefficient (
D
), and pseudo-diffusion coefficient (
D*
), were estimated using a voxel-wise nonlinear regression on DWI series (10
b
-values from 0 to 800 s/mm
2
). The reference standard was histopathological analysis of hepatic inflammation grade, steatosis grade, and fibrosis stage. Intraclass correlation coefficients (ICC), univariate and multivariate correlation analyses, and areas under receiver operating characteristic curves (AUC) were assessed.
Results
Parameters
f
,
D
, and
D*
had ICCs of 0.860, 0.839, and 0.916, respectively. Correlations of
f
,
D
, and
D*
with inflammation grade were
ρ
= − 0.70,
p
< 0.0001;
ρ
= 0.10,
p
= 0.35; and
ρ
= − 0.27,
p
= 0.010, respectively. When adjusting for fibrosis and steatosis, the correlation between
f
and inflammation (
p
< 0.0001) remained, and that between
f
and fibrosis was also significant to a lesser extent (
p
= 0.002). AUCs of
f
,
D
, and
D*
for distinguishing inflammation grades 0 vs. ≥ 1 were 0.84, 0.53, and 0.70; ≤ 1 vs. ≥ 2 were 0.88, 0.57, and 0.60; and ≤ 2 vs. 3 were 0.86, 0.54, and 0.65, respectively.
Conclusion
Perfusion fraction
f
strongly correlated,
D
very weakly correlated, and
D*
weakly correlated with inflammation. Among all IVIM parameters,
f
accurately graded inflammation and showed promise as a biomarker of hepatic inflammation.
Key Points
•
IVIM parameters derived from DWI series with 10 b-values are reproducible for liver tissue characterization
.
•
This retrospective two-center study showed that perfusion fraction provided good diagnostic performance for distinguishing dichotomized grades of inflammation
.
•
Fibrosis is a significant confounder on the association between inflammation and perfusion fraction
.
Purpose
To evaluate the diagnostic performance of Liver Imaging Reporting and Data System (LI-RADS) v2017 major features, the impact of ancillary features, and categories on contrast-enhanced ...computed tomography (CECT) for the diagnosis of hepatocellular carcinoma (HCC).
Materials and methods
This retrospective study included 59 patients (104 observations including 72 HCCs) with clinical suspicion of HCC undergoing CECT between 2013 and 2016. Two radiologists independently assessed major and ancillary imaging features for each liver observation and assigned a LI-RADS category based on major features only and in combination with ancillary features. The composite reference standard included pathology or imaging. Per-lesion estimates of diagnostic performance of major features, ancillary features, and LI-RADS categories were assessed by generalized estimating equation models.
Results
Major features (arterial phase hyperenhancement, washout, capsule, and threshold growth) respectively had a sensitivity of 86.1%, 81.6%, 20.7%, and 26.1% and specificity of 39.3%, 67.9%, 89.9%, and 85.0% for HCC. Ancillary features (ultrasound visibility as discrete nodule, subthreshold growth, and fat in mass more than adjacent liver) respectively had a sensitivity of 42.6%, 50.8%, and 15.1% and a specificity of 79.2%, 66.9%, and 96.4% for HCC. Ancillary features modified the final category in 4 of 104 observations. For HCC diagnosis, categories LR-3, LR-4, LR-5, and LR-TIV (tumor in vein) had a sensitivity of 5.3%, 29.0%, 53.7%, and 10.7%; and a specificity of 49.1%, 84.4%, 97.3%, and 96.4%, respectively.
Conclusion
On CT, LR-5 category has near-perfect specificity for the diagnosis of HCC and ancillary features modifies the final category in few observations.
Purpose
R2* relaxometry is a quantitative method for assessment of iron overload. The purpose is to analyze the cross-sectional relationships between R2* in organs across patients with primary and ...secondary iron overload. Secondary analyses were conducted to analyze R2* according to treatment regimen.
Methods
This is a retrospective, cross-sectional, institutional review board-approved study of eighty-one adult patients with known or suspected iron overload. R2* was measured by segmenting the liver, spleen, bone marrow, pancreas, renal cortex, renal medulla, and myocardium using breath-hold multi-echo gradient-recalled echo imaging at 1.5 T. Phlebotomy, transfusion, and chelation therapy were documented. Analyses included correlation, Kruskal–Wallis, and post hoc Dunn tests.
p
< 0.01 was considered significant.
Results
Correlations between liver R2* and that of the spleen, bone marrow, pancreas, and heart were respectively 0.49, 0.33, 0.27, and 0.34. R2* differed between patients with primary and secondary overload in the liver (
p
< 0.001), spleen (
p
< 0.001), bone marrow (
p
< 0.01), renal cortex (
p
< 0.001), and renal medulla (
p
< 0.001). Liver, spleen, and bone marrow R2* were higher in thalassemia than in hereditary hemochromatosis (all
p
< 0.01). Renal cortex R2* was higher in sickle cell disease than in hereditary hemochromatosis (
p
< 0.001) and in thalassemia (
p
< 0.001). Overall, there was a trend toward lower liver R2* in patients assigned to phlebotomy and higher liver R2* in patients assigned to transfusion and chelation therapy.
Conclusion
R2* relaxometry revealed differences in degree or distribution of iron overload between organs, underlying etiologies, and treatment.
In developed countries, colorectal cancer is the second cause of cancer-related mortality. Chemotherapy is considered a standard treatment for colorectal liver metastases (CLM). Among patients who ...develop CLM, the assessment of patient response to chemotherapy is often required to determine the need for second-line chemotherapy and eligibility for surgery. However, while FOLFOX-based regimens are typically used for CLM treatment, the identification of responsive patients remains elusive. Computer-aided diagnosis systems may provide insight in the classification of liver metastases identified on diagnostic images. In this paper, we propose a fully automated framework based on deep convolutional neural networks (DCNN) which first differentiates treated and untreated lesions to identify new lesions appearing on CT scans, followed by a fully connected neural networks to predict from untreated lesions in pre-treatment computed tomography (CT) for patients with CLM undergoing chemotherapy, their response to a FOLFOX with Bevacizumab regimen as first-line of treatment. The ground truth for assessment of treatment response was histopathology-determined tumor regression grade. Our DCNN approach trained on 444 lesions from 202 patients achieved accuracies of 91% for differentiating treated and untreated lesions, and 78% for predicting the response to FOLFOX-based chemotherapy regimen. Experimental results showed that our method outperformed traditional machine learning algorithms and may allow for the early detection of non-responsive patients.
Introduction
The interest in measuring brain perfusion with intravoxel incoherent motion (IVIM) MRI has significantly increased in the last 3 years. Our aim was to evaluate the prognostic value for ...survival of intravoxel incoherent motion perfusion fraction in patients with gliomas, and compare it to dynamic susceptibility contrast relative cerebral blood volume and apparent diffusion coefficient.
Methods
Images were acquired in 27 patients with brain gliomas (16 high grades, 11 low grades), before any relevant treatment. Region of maximal perfusion fraction, maximal relative cerebral blood volume, and minimal apparent diffusion coefficient were obtained. The accuracy of all three methods for 2‑year survival prognosis was compared using the area under the receiver operating characteristic curve and Kaplan–Meier survival curves.
Results
Death or survival for at least 2 years after imaging could be documented in 22/27 patients. The cutoff values of 0.112 for the perfusion fraction, of 3.01 for the relative cerebral blood volume, and 1033 × 10
−6
mm
2
/s for apparent diffusion coefficient led to an identical sensitivity of 0.889, and a specificity of 0.833, 0.517, and 0.750, respectively for 2 year survival prognosis. The corresponding areas under the receiver operating characteristic curves were 0.84, 076, and 0.86, respectively. All three methods had a significant log rank test considering overall survival (
p
= 0.001,
p
= 0.028, and
p
= 0.002).
Conclusion
In this relatively small cohort, maximal IVIM perfusion fraction, similarly to maximal relative cerebral blood volume and minimal apparent diffusion coefficient, was prognostic for survival in patients with gliomas. Maximal IVIM perfusion fraction and minimal apparent diffusion coefficient performed similarly in predicting survival, and both slightly outperformed maximal relative cerebral blood volume.
Computed tomography (CT) and magnetic resonance imaging (MRI) play critical roles for assessing treatment response of hepatocellular carcinoma (HCC) after locoregional therapy. Interpretation is ...challenging because posttreatment imaging findings depend on the type of treatment, magnitude of treatment response, time interval after treatment, and other factors. To help radiologists interpret and report treatment response in a clear, simple, and standardized manner, the Liver Imaging Reporting and Data System (LI-RADS) has developed a Treatment Response (LR-TR) algorithm. Introduced in 2017, the system provides criteria to categorize response of HCC to locoregional treatment (e.g., chemical ablation, energy-based ablation, transcatheter therapy, and radiation therapy). LR-TR categories include Nonevaluable, Nonviable, Equivocal, and Viable. LR-TR does not apply to patients on systemic therapies. This article reviews the LR-TR algorithm; discusses locoregional therapies for HCC, treatment concepts, and expected posttreatment findings; and illustrates LI-RADS treatment response assessment with CT and MRI.
Abstract Objective To determine if diagnostic signs of adhesive capsulitis (AC) of the shoulder at Magnetic Resonance Imaging (MRI) and arthrography (MRA) are applicable to CT arthrography (CTA). ...Methods 22 shoulder CTAs with AC were retrospectively reviewed for features described in MR literature. The control group was composed of 83 shoulder CTA divided into four subgroups 1) normal (N = 20), 2) omarthrosis (N = 19), 3) labral injury (N = 23), and 4) rotator cuff tear (N = 21). Two musculoskeletal radiologists assessed the rotator interval (RI) for obliteration, increased width and thickening of coracohumeral ligament (CHL). The width and capsule thickness of the axillary recess were measured. Results The width of the axillary recess was significantly decreased in the AC group (4.6 ± 2.6 mm versus 9.9 ± 4.6 mm, p ≤ 0.0001; sensitivity and specificity of 84% and 80%). Thickness of the medial and lateral walls of the axillary capsule was significantly increased in the AC group (5.9 ± 1.3 mm versus 3.7 ± 1.1 mm, p ≤ 0.0001 and 5.7 ± 1 mm versus 3.5 ± 1.3 mm, p ≤ 0.0001, respectively). CHL thickness was significantly increased in the AC group (4.1 ± 1 mm (p ≤ 0.001)) in comparison to others groups. Obliteration of the RI was statistically significantly more frequent in patients with AC (72.7% (16/22) vs. 12% (10/83), p < 0.0001). Width of the RI did not differ significantly between patients and controls (p ≥ 0.428). Conclusion Decreased axillary width, and thickened axillary capsule are MR signs of AC applicable to CTA. Evaluation of rotator interval seems useful and reproducible only for obliteration.