Automatic Detection of Pancreatic Lesions and Main Pancreatic Duct Dilatation on Portal Venous CT Scans Using Deep Learning.

OBJECTIVES: This study proposes and evaluates a deep learning method to detect pancreatic neoplasms and to identify main pancreatic duct (MPD) dilatation on portal venous computed tomography scans. MATERIALS AND METHODS: A total of 2890 portal venous computed tomography scans from 9 institutions were acquired, among which 2185 had a pancreatic neoplasm and 705 were healthy controls. Each scan was reviewed by one in a group of 9 radiologists. Physicians contoured the pancreas, pancreatic lesions if present, and the MPD if visible. They also assessed tumor type and MPD dilatation. Data were split into a training and independent testing set of 2134 and 756 cases, respectively.A method to detect pancreatic lesions and MPD dilatation was built in 3 steps. First, a segmentation network was trained in a 5-fold cross-validation manner. Second, outputs of this network were postprocessed to extract imaging features: a normalized lesion risk, the predicted lesion diameter, and the MPD diameter in the head, body, and tail of the pancreas. Third, 2 logistic regression models were calibrated to predict lesion presence and MPD dilatation, respectively. Performance was assessed on the independent test cohort using receiver operating characteristic analysis. The method was also evaluated on subgroups defined based on lesion types and characteristics. RESULTS: The area under the curve of the model detecting lesion presence in a patient was 0.98 (95% confidence interval [CI], 0.97-0.99). A sensitivity of 0.94 (469 of 493; 95% CI, 0.92-0.97) was reported. Similar values were obtained in patients with small (less than 2 cm) and isodense lesions with a sensitivity of 0.94 (115 of 123; 95% CI, 0.87-0.98) and 0.95 (53 of 56, 95% CI, 0.87-1.0), respectively. The model sensitivity was also comparable across lesion types with values of 0.94 (95% CI, 0.91-0.97), 1.0 (95% CI, 0.98-1.0), 0.96 (95% CI, 0.97-1.0) for pancreatic ductal adenocarcinoma, neuroendocrine tumor, and intraductal papillary neoplasm, respectively. Regarding MPD dilatation detection, the model had an area under the curve of 0.97 (95% CI, 0.96-0.98). CONCLUSIONS: The proposed approach showed high quantitative performance to identify patients with pancreatic neoplasms and to detect MPD dilatation on an independent test cohort. Performance was robust across subgroups of patients with different lesion characteristics and types. Results confirmed the interest to combine a direct lesion detection approach with secondary features such as the MPD diameter, thus indicating a promising avenue for the detection of pancreatic cancer at early stages.

From Dose Reduction to Contrast Maximization: Can Deep Learning Amplify the Impact of Contrast Media on Brain Magnetic Resonance Image Quality? A Reader Study.

OBJECTIVES: The aim of this study was to evaluate a deep learning method designed to increase the contrast-to-noise ratio in contrast-enhanced gradient echo T1-weighted brain magnetic resonance imaging (MRI) acquisitions. The processed images are quantitatively evaluated in terms of lesion detection performance. MATERIALS AND METHODS: A total of 250 multiparametric brain MRIs, acquired between November 2019 and March 2021 at Gustave Roussy Cancer Campus (Villejuif, France), were considered for inclusion in this retrospective monocentric study. Independent training (107 cases; age, 55 ± 14 years; 58 women) and test (79 cases; age, 59 ± 14 years; 41 women) samples were defined. Patients had glioma, brain metastasis, meningioma, or no enhancing lesion. Gradient echo and turbo spin echo with variable flip angles postcontrast T1 sequences were acquired in all cases. For the cases that formed the training sample, "low-dose" postcontrast gradient echo T1 images using 0.025 mmol/kg injections of contrast agent were also acquired. A deep neural network was trained to synthetically enhance the low-dose T1 acquisitions, taking standard-dose T1 MRI as reference. Once trained, the contrast enhancement network was used to process the test gradient echo T1 images. A read was then performed by 2 experienced neuroradiologists to evaluate the original and processed T1 MRI sequences in terms of contrast enhancement and lesion detection performance, taking the turbo spin echo sequences as reference. RESULTS: The processed images were superior to the original gradient echo and reference turbo spin echo T1 sequences in terms of contrast-to-noise ratio (44.5 vs 9.1 and 16.8; P < 0.001), lesion-to-brain ratio (1.66 vs 1.31 and 1.44; P < 0.001), and contrast enhancement percentage (112.4% vs 85.6% and 92.2%; P < 0.001) for cases with enhancing lesions. The overall image quality of processed T1 was preferred by both readers (graded 3.4/4 on average vs 2.7/4; P < 0.001). Finally, the proposed processing improved the average sensitivity of gradient echo T1 MRI from 88% to 96% for lesions larger than 10 mm ( P = 0.008), whereas no difference was found in terms of the false detection rate (0.02 per case in both cases; P > 0.99). The same effect was observed when considering all lesions larger than 5 mm: sensitivity increased from 70% to 85% ( P < 0.001), whereas false detection rates remained similar (0.04 vs 0.06 per case; P = 0.48). With all lesions included regardless of their size, sensitivities were 59% and 75% for original and processed T1 images, respectively ( P < 0.001), and the corresponding false detection rates were 0.05 and 0.14 per case, respectively ( P = 0.06). CONCLUSION: The proposed deep learning method successfully amplified the beneficial effects of contrast agent injection on gradient echo T1 image quality, contrast level, and lesion detection performance. In particular, the sensitivity of the MRI sequence was improved by up to 16%, whereas the false detection rate remained similar.

Can Deep Learning Replace Gadolinium in Neuro-Oncology?: A Reader Study.

MATERIALS AND METHODS: This monocentric retrospective study leveraged 200 multiparametric brain MRIs acquired between November 2019 and February 2020 at Gustave Roussy Cancer Campus (Villejuif, France). A total of 145 patients were included: 107 formed the training sample (55 ± 14 years, 58 women) and 38 the separate test sample (62 ± 12 years, 22 women). Patients had glioma, brain metastases, meningioma, or no enhancing lesion. T1, T2-FLAIR, diffusion-weighted imaging, low-dose, and standard-dose postcontrast T1 sequences were acquired. A deep network was trained to process the precontrast and low-dose sequences to predict "virtual" surrogate images for contrast-enhanced T1. Once trained, the deep learning method was evaluated on the test sample. The discrepancies between the predicted virtual images and the standard-dose MRIs were qualitatively and quantitatively evaluated using both automated voxel-wise metrics and a reader study, where 2 radiologists graded image qualities and marked all visible enhancing lesions. RESULTS: The automated analysis of the test brain MRIs computed a structural similarity index of 87.1% ± 4.8% between the predicted virtual sequences and the reference contrast-enhanced T1 MRIs, a peak signal-to-noise ratio of 31.6 ± 2.0 dB, and an area under the curve of 96.4% ± 3.1%. At Youden's operating point, the voxel-wise sensitivity (SE) and specificity were 96.4% and 94.8%, respectively. The reader study found that virtual images were preferred to standard-dose MRI in terms of image quality (P = 0.008). A total of 91 reference lesions were identified in the 38 test T1 sequences enhanced with full dose of contrast agent. On average across readers, the brain lesion SE of the virtual images was 83% for lesions larger than 10 mm (n = 42), and the associated false detection rate was 0.08 lesion/patient. The corresponding positive predictive value of detected lesions was 92%, and the F1 score was 88%. Lesion detection performance, however, dropped when smaller lesions were included: average SE was 67% for lesions larger than 5 mm (n = 74), and 56% with all lesions included regardless of their size. The false detection rate remained below 0.50 lesion/patient in all cases, and the positive predictive value remained above 73%. The composite F1 score was 63% at worst. CONCLUSIONS: The proposed deep learning method for virtual contrast-enhanced T1 brain MRI prediction showed very high quantitative performance when evaluated with standard voxel-wise metrics. The reader study demonstrated that, for lesions larger than 10 mm, good detection performance could be maintained despite a 4-fold division in contrast agent usage, unveiling a promising avenue for reducing the gadolinium exposure of returning patients. Small lesions proved, however, difficult to handle for the deep network, showing that full-dose injections remain essential for accurate first-line diagnosis in neuro-oncology.

Deep Learning Techniques for Automatic MRI Cardiac Multi-Structures Segmentation and Diagnosis: Is the Problem Solved?

Delineation of the left ventricular cavity, myocardium, and right ventricle from cardiac magnetic resonance images (multi-slice 2-D cine MRI) is a common clinical task to establish diagnosis. The automation of the corresponding tasks has thus been the subject of intense research over the past decades. In this paper, we introduce the "Automatic Cardiac Diagnosis Challenge" dataset (ACDC), the largest publicly available and fully annotated dataset for the purpose of cardiac MRI (CMR) assessment. The dataset contains data from 150 multi-equipments CMRI recordings with reference measurements and classification from two medical experts. The overarching objective of this paper is to measure how far state-of-the-art deep learning methods can go at assessing CMRI, i.e., segmenting the myocardium and the two ventricles as well as classifying pathologies. In the wake of the 2017 MICCAI-ACDC challenge, we report results from deep learning methods provided by nine research groups for the segmentation task and four groups for the classification task. Results show that the best methods faithfully reproduce the expert analysis, leading to a mean value of 0.97 correlation score for the automatic extraction of clinical indices and an accuracy of 0.96 for automatic diagnosis. These results clearly open the door to highly accurate and fully automatic analysis of cardiac CMRI. We also identify scenarios for which deep learning methods are still failing. Both the dataset and detailed results are publicly available online, while the platform will remain open for new submissions.

Low-dimensional representation of cardiac motion using Barycentric Subspaces: A new group-wise paradigm for estimation, analysis, and reconstruction.

One major challenge when trying to build low-dimensional representation of the cardiac motion is its natural circular pattern during a cycle, therefore making the mean image a poor descriptor of the whole sequence. Therefore, traditional approaches for the analysis of the cardiac deformation use one specific frame of the sequence - the end-diastolic (ED) frame - as a reference to study the whole motion. Consequently, this methodology is biased by this empirical choice. Moreover, the ED image might be a poor reference when looking at large deformation for example at the end-systolic (ES) frame. In this paper, we propose a novel approach to study cardiac motion in 4D image sequences using low-dimensional subspace analysis. Instead of building subspaces relying on a mean value we use a novel type of subspaces called Barycentric Subspaces which are implicitly defined as the weighted Karcher means of k+1 reference images instead of being defined with respect to one reference image. In the first part of this article, we introduce the methodological framework and the algorithms used to manipulate images within these new subspaces: how to compute the projection of a given image on the Barycentric Subspace with its coordinates, and the opposite operation of computing an image from a set of references and coordinates. Then we show how this framework can be applied to cardiac motion problems and lead to significant improvements over the single reference method. Firstly, by computing the low-dimensional representation of two populations we show that the parameters extracted correspond to relevant cardiac motion features leading to an efficient representation and discrimination of both groups. Secondly, in motion estimation, we use the projection on this low-dimensional subspace as an additional prior on the regularization in cardiac motion tracking, efficiently reducing the error of the registration between the ED and ES by almost 30%. We also derive a symmetric and transitive formulation of the registration that can be used both for frame-to-frame and frame-to-reference registration. Finally, we look at the reconstruction of the images using our proposed low-dimensional representation and show that this multi-references method using Barycentric Subspaces performs better than traditional approaches based on a single reference.

Statistical shape modeling of the left ventricle: myocardial infarct classification challenge.

Statistical shape modeling is a powerful tool for visualizing and quantifying geometric and functional patterns of the heart. After myocardial infarction (MI), the left ventricle typically remodels in response to physiological challenges. Several methods have been proposed in the literature to describe statistical shape changes. Which method best characterizes left ventricular remodeling after MI is an open research question. A better descriptor of remodeling is expected to provide a more accurate evaluation of disease status in MI patients. We therefore designed a challenge to test shape characterization in MI given a set of three-dimensional left ventricular surface points. The training set comprised 100 MI patients, and 100 asymptomatic volunteers (AV). The challenge was initiated in 2015 at the Statistical Atlases and Computational Models of the Heart workshop, in conjunction with the MICCAI conference. The training set with labels was provided to participants, who were asked to submit the likelihood of MI from a different (validation) set of 200 cases (100 AV and 100 MI). Sensitivity, specificity, accuracy and area under the receiver operating characteristic curve were used as the outcome measures. The goals of this challenge were to (1) establish a common dataset for evaluating statistical shape modeling algorithms in MI, and (2) test whether statistical shape modeling provides additional information characterizing MI patients over standard clinical measures. Eleven groups with a wide variety of classification and feature extraction approaches participated in this challenge. All methods achieved excellent classification results with accuracy ranges from 0.83 to 0.98. The areas under the receiver operating characteristic curves were all above 0.90. Four methods showed significantly higher performance than standard clinical measures. The dataset and software for evaluation are available from the Cardiac Atlas Project website1.