Prostate focal peripheral zone lesions: characterization at multiparametric MR imaging--influence of a computer-aided diagnosis system.

PURPOSE: To assess the impact of a computer-aided diagnosis (CAD) system in the characterization of focal prostate lesions at multiparametric magnetic resonance (MR) imaging. MATERIALS AND METHODS: Formal institutional review board approval was not required. Thirty consecutive 1.5-T multiparametric MR imaging studies (with T2-weighted, diffusion-weighted, and dynamic contrast material-enhanced imaging) obtained before radical prostatectomy in patients between September 2008 and February 2010 were reviewed. Twelve readers assessed the likelihood of malignancy of 88 predefined peripheral zone lesions by using a five-level (level, 0-4) subjective score (SS) in reading session 1. This was repeated 5 weeks later in reading session 2. The CAD results were then disclosed, and in reading session 3, the readers could amend the scores assigned during reading session 2. Diagnostic accuracy was assessed by using a receiver operating characteristic (ROC) regression model and was quantified with the area under the ROC curve (AUC). RESULTS: Mean AUCs were significantly lower for less experienced (<1 year) readers (P < .02 for all sessions). Seven readers improved their performance between reading sessions 1 and 2, and 12 readers improved their performance between sessions 2 and 3. The mean AUCs for reading session 1 (83.0%; 95% confidence interval [CI]: 77.9%, 88.0%) and reading session 2 (84.1%; 95% CI: 78.1%, 88.7%) were not significantly different (P = .76). Although the mean AUC for reading session 3 (87.2%; 95% CI: 81.0%, 92.0%) was higher than that for session 2, the difference was not significant (P = .08). For an SS positivity threshold of 3, the specificity of reading session 2 (79.0%; 95% CI: 71.1%, 86.4%) was not significantly different from that of session 1 (78.7%; 95% CI: 70.5%, 86.8%) but was significantly lower than that of session 3 (86.2%; 95% CI: 77.1%, 93.1%; P < .03). The sensitivity of reading session 2 (68.4%; 95% CI: 57.5%, 77.7%) was significantly higher than that of session 1 (64.0%; 95% CI: 52.9%, 73.9%; P = .003) but was not significantly different from that of session 3 (71.4%; 95% CI: 58.3%, 82.7%). CONCLUSION: A CAD system may improve the characterization of prostate lesions at multiparametric MR imaging by increasing reading specificity.

OntoVIP: an ontology for the annotation of object models used for medical image simulation.

This paper describes the creation of a comprehensive conceptualization of object models used in medical image simulation, suitable for major imaging modalities and simulators. The goal is to create an application ontology that can be used to annotate the models in a repository integrated in the Virtual Imaging Platform (VIP), to facilitate their sharing and reuse. Annotations make the anatomical, physiological and pathophysiological content of the object models explicit. In such an interdisciplinary context we chose to rely on a common integration framework provided by a foundational ontology, that facilitates the consistent integration of the various modules extracted from several existing ontologies, i.e. FMA, PATO, MPATH, RadLex and ChEBI. Emphasis is put on methodology for achieving this extraction and integration. The most salient aspects of the ontology are presented, especially the organization in model layers, as well as its use to browse and query the model repository.

Federated brain tumor segmentation: An extensive benchmark.

Recently, federated learning has raised increasing interest in the medical image analysis field due to its ability to aggregate multi-center data with privacy-preserving properties. A large amount of federated training schemes have been published, which we categorize into global (one final model), personalized (one model per institution) or hybrid (one model per cluster of institutions) methods. However, their applicability on the recently published Federated Brain Tumor Segmentation 2022 dataset has not been explored yet. We propose an extensive benchmark of federated learning algorithms from all three classes on this task. While standard FedAvg already performs very well, we show that some methods from each category can bring a slight performance improvement and potentially limit the final model(s) bias toward the predominant data distribution of the federation. Moreover, we provide a deeper understanding of the behavior of federated learning on this task through alternative ways of distributing the pooled dataset among institutions, namely an Independent and Identical Distributed (IID) setup, and a limited data setup. Our code is available at (https://github.com/MatthisManthe/Benchmark_FeTS2022).

Deep-learning predicted PET can be subtracted from the true clinical fluorodeoxyglucose PET co-registered to MRI to identify the epileptogenic zone in focal epilepsy.

OBJECTIVE: Normal interictal [18 F]FDG-PET can be predicted from the corresponding T1w MRI with Generative Adversarial Networks (GANs). A technique we call SIPCOM (Subtraction Interictal PET Co-registered to MRI) can then be used to compare epilepsy patients' predicted and clinical PET. We assessed the ability of SIPCOM to identify the Resection Zone (RZ) in patients with drug-resistant epilepsy (DRE) with reference to visual and statistical parametric mapping (SPM) analysis. METHODS: Patients with complete presurgical work-up and subsequent SEEG and cortectomy were included. RZ localisation, the reference region, was assigned to one of eighteen anatomical brain regions. SIPCOM was implemented using healthy controls to train a GAN. To compare, the clinical PET coregistered to MRI was visually assessed by two trained readers, and a standard SPM analysis was performed. RESULTS: Twenty patients aged 17-50 (32 ± 7.8) years were included, 14 (70%) with temporal lobe epilepsy (TLE). Eight (40%) were MRI-negative. After surgery, 14 patients (70%) had a good outcome (Engel I-II). RZ localisation rate was 60% with SIPCOM vs 35% using SPM (P = 0.015) and vs 85% using visual analysis (P = 0.54). Results were similar for Engel I-II patients, the RZ localisation rate was 64% with SIPCOM vs 36% with SPM. With SIPCOM localisation was correct in 67% in MRI-positive vs 50% in MRI-negative patients, and 64% in TLE vs 43% in extra-TLE. The average number of false-positive clusters was 2.2 ± 1.3 using SIPCOM vs 2.3 ± 3.1 using SPM. All RZs localized with SPM were correctly localized with SIPCOM. In one case, PET and MRI were visually reported as negative, but both SIPCOM and SPM localized the RZ. SIGNIFICANCE: SIPCOM performed better than the reference computer-assisted method (SPM) for RZ detection in a group of operated DRE patients. SIPCOM's impact on epilepsy management needs to be prospectively validated.

ProstAttention-Net: A deep attention model for prostate cancer segmentation by aggressiveness in MRI scans.

Multiparametric magnetic resonance imaging (mp-MRI) has shown excellent results in the detection of prostate cancer (PCa). However, characterizing prostate lesions aggressiveness in mp-MRI sequences is impossible in clinical practice, and biopsy remains the reference to determine the Gleason score (GS). In this work, we propose a novel end-to-end multi-class network that jointly segments the prostate gland and cancer lesions with GS group grading. After encoding the information on a latent space, the network is separated in two branches: 1) the first branch performs prostate segmentation 2) the second branch uses this zonal prior as an attention gate for the detection and grading of prostate lesions. The model was trained and validated with a 5-fold cross-validation on a heterogeneous series of 219 MRI exams acquired on three different scanners prior prostatectomy. In the free-response receiver operating characteristics (FROC) analysis for clinically significant lesions (defined as GS >6) detection, our model achieves 69.0%±14.5% sensitivity at 2.9 false positive per patient on the whole prostate and  70.8%±14.4% sensitivity at 1.5 false positive when considering the peripheral zone (PZ) only. Regarding the automatic GS group grading, Cohen's quadratic weighted kappa coefficient (κ) is 0.418±0.138, which is the best reported lesion-wise kappa for GS segmentation to our knowledge. The model has encouraging generalization capacities with κ=0.120±0.092 on the PROSTATEx-2 public dataset and achieves state-of-the-art performance for the segmentation of the whole prostate gland with a Dice of 0.875±0.013. Finally, we show that ProstAttention-Net improves performance in comparison to reference segmentation models, including U-Net, DeepLabv3+ and E-Net. The proposed attention mechanism is also shown to outperform Attention U-Net.

PET image enhancement using artificial intelligence for better characterization of epilepsy lesions.

Introduction: [18F]fluorodeoxyglucose ([18F]FDG) brain PET is used clinically to detect small areas of decreased uptake associated with epileptogenic lesions, e.g., Focal Cortical Dysplasias (FCD) but its performance is limited due to spatial resolution and low contrast. We aimed to develop a deep learning-based PET image enhancement method using simulated PET to improve lesion visualization. Methods: We created 210 numerical brain phantoms (MRI segmented into 9 regions) and assigned 10 different plausible activity values (e.g., GM/WM ratios) resulting in 2100 ground truth high quality (GT-HQ) PET phantoms. With a validated Monte-Carlo PET simulator, we then created 2100 simulated standard quality (S-SQ) [18F]FDG scans. We trained a ResNet on 80% of this dataset (10% used for validation) to learn the mapping between S-SQ and GT-HQ PET, outputting a predicted HQ (P-HQ) PET. For the remaining 10%, we assessed Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Root Mean Squared Error (RMSE) against GT-HQ PET. For GM and WM, we computed recovery coefficients (RC) and coefficient of variation (COV). We also created lesioned GT-HQ phantoms, S-SQ PET and P-HQ PET with simulated small hypometabolic lesions characteristic of FCDs. We evaluated lesion detectability on S-SQ and P-HQ PET both visually and measuring the Relative Lesion Activity (RLA, measured activity in the reduced-activity ROI over the standard-activity ROI). Lastly, we applied our previously trained ResNet on 10 clinical epilepsy PETs to predict the corresponding HQ-PET and assessed image quality and confidence metrics. Results: Compared to S-SQ PET, P-HQ PET improved PNSR, SSIM and RMSE; significatively improved GM RCs (from 0.29 ± 0.03 to 0.79 ± 0.04) and WM RCs (from 0.49 ± 0.03 to 1 ± 0.05); mean COVs were not statistically different. Visual lesion detection improved from 38 to 75%, with average RLA decreasing from 0.83 ± 0.08 to 0.67 ± 0.14. Visual quality of P-HQ clinical PET improved as well as reader confidence. Conclusion: P-HQ PET showed improved image quality compared to S-SQ PET across several objective quantitative metrics and increased detectability of simulated lesions. In addition, the model generalized to clinical data. Further evaluation is required to study generalization of our method and to assess clinical performance in larger cohorts.

Detection of ISUP ≥2 prostate cancers using multiparametric MRI: prospective multicentre assessment of the non-inferiority of an artificial intelligence system as compared to the PI-RADS V.2.1 score (CHANGE study).

INTRODUCTION: Prostate multiparametric MRI (mpMRI) has shown good sensitivity in detecting cancers with an International Society of Urological Pathology (ISUP) grade of ≥2. However, it lacks specificity, and its inter-reader reproducibility remains moderate. Biomarkers, such as the Prostate Health Index (PHI), may help select patients for prostate biopsy. Computer-aided diagnosis/detection (CAD) systems may also improve mpMRI interpretation. Different prototypes of CAD systems are currently developed under the Recherche Hospitalo-Universitaire en Santé / Personalized Focused Ultrasound Surgery of Localized Prostate Cancer (RHU PERFUSE) research programme, tackling challenging issues such as robustness across imaging protocols and magnetic resonance (MR) vendors, and ability to characterise cancer aggressiveness. The study primary objective is to evaluate the non-inferiority of the area under the receiver operating characteristic curve of the final CAD system as compared with the Prostate Imaging-Reporting and Data System V.2.1 (PI-RADS V.2.1) in predicting the presence of ISUP ≥2 prostate cancer in patients undergoing prostate biopsy. METHODS: This prospective, multicentre, non-inferiority trial will include 420 men with suspected prostate cancer, a prostate-specific antigen level of ≤30 ng/mL and a clinical stage ≤T2 c. Included men will undergo prostate mpMRI that will be interpreted using the PI-RADS V.2.1 score. Then, they will undergo systematic and targeted biopsy. PHI will be assessed before biopsy. At the end of patient inclusion, MR images will be assessed by the final version of the CAD system developed under the RHU PERFUSE programme. Key secondary outcomes include the prediction of ISUP grade ≥2 prostate cancer during a 3-year follow-up, and the number of biopsy procedures saved and ISUP grade ≥2 cancers missed by several diagnostic pathways combining PHI and MRI findings. ETHICS AND DISSEMINATION: Ethical approval was obtained from the Comité de Protection des Personnes Nord Ouest III (ID-RCB: 2020-A02785-34). After publication of the results, access to MR images will be possible for testing other CAD systems. TRIAL REGISTRATION NUMBER: NCT04732156.

Regularized siamese neural network for unsupervised outlier detection on brain multiparametric magnetic resonance imaging: Application to epilepsy lesion screening.

In this study, we propose a novel anomaly detection model targeting subtle brain lesions in multiparametric MRI. To compensate for the lack of annotated data adequately sampling the heterogeneity of such pathologies, we cast this problem as an outlier detection problem and introduce a novel configuration of unsupervised deep siamese networks to learn normal brain representations using a series of non-pathological brain scans. The proposed siamese network, composed of stacked convolutional autoencoders as subnetworks is designed to map patches extracted from healthy control scans only and centered at the same spatial localization to 'close' representations with respect to the chosen metric in a latent space. It is based on a novel loss function combining a similarity term and a regularization term compensating for the lack of dissimilar pairs. These latent representations are then fed into oc-SVM models at voxel-level to produce anomaly score maps. We evaluate the performance of our brain anomaly detection model to detect subtle epilepsy lesions in multiparametric (T1-weighted, FLAIR) MRI exams considered as normal (MRI-negative). Our detection model trained on 75 healthy subjects and validated on 21 epilepsy patients (with 18 MRI-negatives) achieves a maximum sensitivity of 61% on the MRI-negative lesions, identified among the 5 most suspicious detections on average. It is shown to outperform detection models based on the same architecture but with stacked convolutional or Wasserstein autoencoders as unsupervised feature extraction mechanisms.

LU-Net: A Multistage Attention Network to Improve the Robustness of Segmentation of Left Ventricular Structures in 2-D Echocardiography.

Segmentation of cardiac structures is one of the fundamental steps to estimate volumetric indices of the heart. This step is still performed semiautomatically in clinical routine and is, thus, prone to interobserver and intraobserver variabilities. Recent studies have shown that deep learning has the potential to perform fully automatic segmentation. However, the current best solutions still suffer from a lack of robustness in terms of accuracy and number of outliers. The goal of this work is to introduce a novel network designed to improve the overall segmentation accuracy of left ventricular structures (endocardial and epicardial borders) while enhancing the estimation of the corresponding clinical indices and reducing the number of outliers. This network is based on a multistage framework where both the localization and segmentation steps are optimized jointly through an end-to-end scheme. Results obtained on a large open access data set show that our method outperforms the current best-performing deep learning solution with a lighter architecture and achieved an overall segmentation accuracy lower than the intraobserver variability for the epicardial border (i.e., on average a mean absolute error of 1.5 mm and a Hausdorff distance of 5.1mm) with 11% of outliers. Moreover, we demonstrate that our method can closely reproduce the expert analysis for the end-diastolic and end-systolic left ventricular volumes, with a mean correlation of 0.96 and a mean absolute error of 7.6 ml. Concerning the ejection fraction of the left ventricle, results are more contrasted with a mean correlation coefficient of 0.83 and an absolute mean error of 5.0%, producing scores that are slightly below the intraobserver margin. Based on this observation, areas for improvement are suggested.

Deep Learning for Segmentation Using an Open Large-Scale Dataset in 2D Echocardiography.

Delineation of the cardiac structures from 2D echocardiographic images is a common clinical task to establish a diagnosis. Over the past decades, the automation of this task has been the subject of intense research. In this paper, we evaluate how far the state-of-the-art encoder-decoder deep convolutional neural network methods can go at assessing 2D echocardiographic images, i.e., segmenting cardiac structures and estimating clinical indices, on a dataset, especially, designed to answer this objective. We, therefore, introduce the cardiac acquisitions for multi-structure ultrasound segmentation dataset, the largest publicly-available and fully-annotated dataset for the purpose of echocardiographic assessment. The dataset contains two and four-chamber acquisitions from 500 patients with reference measurements from one cardiologist on the full dataset and from three cardiologists on a fold of 50 patients. Results show that encoder-decoder-based architectures outperform state-of-the-art non-deep learning methods and faithfully reproduce the expert analysis for the end-diastolic and end-systolic left ventricular volumes, with a mean correlation of 0.95 and an absolute mean error of 9.5 ml. Concerning the ejection fraction of the left ventricle, results are more contrasted with a mean correlation coefficient of 0.80 and an absolute mean error of 5.6%. Although these results are below the inter-observer scores, they remain slightly worse than the intra-observer's ones. Based on this observation, areas for improvement are defined, which open the door for accurate and fully-automatic analysis of 2D echocardiographic images.

Detection of Lesions Underlying Intractable Epilepsy on T1-Weighted MRI as an Outlier Detection Problem.

Pattern recognition methods, such as computer aided diagnosis (CAD) systems, can help clinicians in their diagnosis by marking abnormal regions in an image. We propose a machine learning system based on a one-class support vector machine (OC-SVM) classifier for the detection of abnormalities in magnetic resonance images (MRI) applied to patients with intractable epilepsy. The system learns the features associated with healthy control subjects, allowing a voxelwise assessment of the deviation of a test subject pattern from the learned patterns. While any number of various features can be chosen and learned, here we focus on two texture parameters capturing image patterns associated with epileptogenic lesions on T1-weighted brain MRI e.g. heterotopia and blurred junction between the grey and white matter. The CAD output consists of patient specific 3D maps locating clusters of suspicious voxels ranked by size and degree of deviation from control patterns. System performance was evaluated using realistic simulations of challenging detection tasks as well as clinical data of 77 healthy control subjects and of eleven patients (13 lesions). It was compared to that of a mass univariate statistical parametric mapping (SPM) single subject analysis based on the same set of features. For all simulations, OC-SVM yielded significantly higher values of the area under the ROC curve (AUC) and higher sensitivity at low false positive rate. For the clinical data, both OC-SVM and SPM successfully detected 100% of the lesions in the MRI positive cases (3/13). For the MRI negative cases (10/13), OC-SVM detected 7/10 lesions and SPM analysis detected 5/10 lesions. In all experiments, OC-SVM produced fewer false positive detections than SPM. OC-SVM may be a versatile system for unbiased lesion detection.

Correction methods for random coincidences in fully 3D whole-body PET: impact on data and image quality.

UNLABELLED: With the advantages of the increased sensitivity of fully 3-dimensional (3D) PET for whole-body imaging come the challenges of more complicated quantitative corrections and, in particular, an increase in the number of random coincidences. The most common method of correcting for random coincidences is the real-time subtraction of a delayed coincidence channel, which does not add bias but increases noise. An alternative approach is the postacquisition subtraction of a low-noise random coincidence estimate, which can be obtained either from a smoothed delayed coincidence sinogram or from a calibration scan or directly estimated. Each method makes different trade-offs between noise amplification, bias, and data-processing requirements. These trade-offs are dependent on activity injected, the local imaging environment (e.g., near the bladder), and the reconstruction algorithm. METHODS: Using fully 3D whole-body simulations and phantom studies, we investigate how the gains in noise equivalent count (NEC) rates from using a noiseless random coincidence estimation method are translated to improvements in image signal-to-noise ratio (SNR). The image SNR, however, depends on the image reconstruction method and the local imaging environment. RESULTS: We show that for fully 3D whole-body imaging using a particular set of scanners and clinical protocols, a low-noise estimate of random coincidences improves sinogram and image SNRs by approximately 15% compared with online subtraction of delayed coincidences. CONCLUSION: A 15% improvement in image SNR arises from a 32% increase in the NEC rate. Thus, scan duration can be reduced by 25% while still maintaining a constant total acquired NEC.

A lesion detection observer study comparing 2-dimensional versus fully 3-dimensional whole-body PET imaging protocols.

UNLABELLED: We compared the impact of 2-dimensional (2D) and fully 3-dimensional (3D) acquisition modes on the performance of human observers in detecting and localizing tumors in whole-body (18)F-FDG images. METHODS: We selected protocols based on noise equivalent count (NEC) rates derived from a series of 2D and fully 3D whole-body patient and phantom acquisitions on a dual-mode PET scanner. The fully 3D peak NEC value for a standard 70-kg patient was achieved for an injected dose of approximately 444 MBq (12 mCi) assuming a 90-min delay before acquisition, whereas the 2D peak value was never reached. The protocols were therefore set to those corresponding to a 444-MBq injected dose in fully 3D and 2D and a 740-MBq (20 mCi) injected dose in 2D that was considered as the maximum allowable dose. We used a non-Monte Carlo simulator to generate multiple realizations of whole-body PET data based on the geometry of the mathematic cardiac torso phantom (MCAT) with accurate noise properties. Two-dimensional and fully 3D acquisition times were set to 5 min per bed position. Spherical 1-cm-diameter lesions (targets) with random locations and contrasts were distributed in different organs. The simulated 2D datasets were reconstructed using attenuation-weighted ordered-subsets expectation maximization ((AW)OSEM) and the fully 3D datasets were reconstructed with FORE+(AW)OSEM (FORE = Fourier rebinning). Five human observers located and ranked the targets using a volumetric display of the whole-body PET data to replicate the clinical practice. An alternate free-response operating characteristic (AFROC) analysis of the human observer reports was performed for each protocol and each organ separately. RESULTS: The 2D protocol corresponding to 740-MBq injected dose allowed the overall best detection performance. It was followed by the fully 3D acquisition at the peak fully 3D NEC rate from a 444-MBq injected dose. A 2D acquisition corresponding to a 444-MBq injected dose was ranked last. Differences in detection performance were organ specific. CONCLUSION: This study showed that, for this patient size and scanner type, the fully 3D acquisition mode allowed better or equivalent detection performance than the 2D mode for an injected dose corresponding to the peak fully 3D NEC rate. The 2D acquisition protocol combined with a higher injected dose resulted in the highest detectabilities.

Optimization of injected dose based on noise equivalent count rates for 2- and 3-dimensional whole-body PET.

UNLABELLED: The noise equivalent count (NEC) rate index is used to derive guidelines on the optimal injected dose to the patient for 2-dimensional (2D) and 3-dimensional (3D) whole-body PET acquisitions. METHODS: We performed 2D and 3D whole-body acquisitions of an anthropomorphic phantom modeling the conditions for (18)F-FDG PET of the torso and measured the NEC rates for different activity levels for several organs of interest. The correlations between count rates measured from the phantom and those from a series of whole-body patient scans were then analyzed. This analysis allowed validation of our approach and estimation of the injected dose that maximizes NEC rate as a function of patient morphology for both acquisition modes. RESULTS: Variations of the phantom and patient prompt and random coincidence rates as a function of single-photon rates correlated well. On the basis of these correlations, we demonstrated that the patient NEC rate can be predicted for a given single-photon rate. Finally, we determined that patient single-photon rates correlated with the mean dose per weight at acquisition start when normalized by the body mass index. This correlation allows modifying the injected dose as a function of patient body mass index to reach the peak NEC rate in 3D mode. Conversely, we found that the peak NEC rates were never reached in 2D mode within an acceptable range of injected dose. CONCLUSION: The injected dose was adapted to patient morphology for 2D and 3D whole-body acquisitions using the NEC rate as a figure of merit of the statistical quality of the sinogram data. This study is a first step toward a more comprehensive comparison of the image quality obtained using both acquisition modes.

Evaluating image reconstruction methods for tumor detection in 3-dimensional whole-body PET oncology imaging.

UNLABELLED: We compare 3 image reconstruction algorithms for use in 3-dimensional (3D) whole-body PET oncology imaging. We have previously shown that combining Fourier rebinning (FORE) with 2-dimensional (2D) statistical image reconstruction via the ordered-subsets expectation-maximization (OSEM) and attenuation-weighted OSEM (AWOSEM) algorithms demonstrates improvements in image signal-to-noise ratios compared with the commonly used analytic 3D reprojection (3DRP) or FORE+FBP (2D filtered backprojection) reconstruction methods. To assess the impact of these reconstruction methods on detecting and localizing small lesions, we performed a human observer study comparing the different reconstruction methods. The observer study used the same volumetric visualization software tool that is used in clinical practice, instead of a planar viewing mode as is generally used with the standard receiver operating characteristic (ROC) methodology. This change in the human evaluation strategy disallowed the use of a ROC analysis, so instead we compared the fraction of actual targets found and reported (fraction-found) and also investigated the use of an alternative free-response operating characteristic (AFROC) analysis. METHODS: We used a non-Monte Carlo technique to generate 50 statistically accurate realizations of 3D whole-body PET data based on an extended mathematic cardiac torso (MCAT) phantom and with noise levels typical of clinical scans performed on a PET scanner. To each realization, we added 7 randomly located 1-cm-diameter lesions (targets) whose contrasts were varied to sample the range of detectability. These targets were inserted in 3 organs of interest: lungs, liver, and soft tissues. The images were reconstructed with 3 reconstruction strategies (FORE+OSEM, FORE+AWOSEM, and FORE+FBP). Five human observers reported (localized and rated) 7 targets within each volume image. An observer's performance accuracy with each algorithm was measured, as a function of the lesion contrast and organ type, by the fraction of those targets reported and by the area below the AFROC curve. This AFROC curve plots the fraction of reported targets at each rating threshold against the fraction of cases with (> or =1) similarly rated false reports. RESULTS: Images reconstructed with FORE+AWOSEM yielded the best overall target detection as compared with FORE+FBP and FORE+OSEM, although these differences in detectability were region specific. The FORE+FBP and FORE+AWOSEM algorithms had similar performances for liver targets. The FORE+OSEM algorithm performed significantly worse at target detection, especially in the liver. We speculate that this is the result of using an incorrect statistical model for OSEM and that the incorporation of attenuation weighting in AWOSEM largely compensates for this model inaccuracy. These results were consistent for both the fraction of actual targets found and the AFROC analysis. CONCLUSION: We demonstrated the efficacy of performing observer detection studies using the same visualization tools as those used in clinical PET oncology imaging. These studies demonstrated that the FORE+AWOSEM algorithm led to the best overall detection and localization performance for 1-cm-diameter targets compared with the FORE+OSEM and FORE+FBP algorithms.

Volumetric model and human observer comparisons of tumor detection for whole-body positron emission tomography.

RATIONALE AND OBJECTIVES: The performances of model and human observers were compared for detecting and localizing tumors in whole-body positron emission tomography imaging. Volumetric implementations of model observers were used instead of standard planar implementations to reproduce the common practice of using volumetric image displays to assess whole-body positron emission tomography images. MATERIALS AND METHODS: Observer studies with simulated data were used to compare three different acquisition protocols for an average patient size. Multiple realizations of simulated whole-body data with multiple added 1-cm diameter spherical lesions (targets) per image volume for efficiency were used. The location and contrast ratio of the targets were chosen randomly within ranges determined by initial calibration studies. Human observer studies were performed using a volumetric image display routinely used in clinical practice, and human observer detection performance was quantified using an alternate free-response operating characteristic analysis. The human detection performances were compared with the performances of volumetric implementations of the non-prewhitening matched filter and the channelized hotelling observer and also to the target contrast measured in the reconstructed image. RESULTS: Human observer detectability was generally well described as a linear function of these three figures of merit for the detection task considered. The best correlations (r = 0.96, rho = 0.98) were achieved with the channelized hotelling observer and non-prewhitening matched filter model observers. CONCLUSION: The use of volumetric model observers provides a means for quantitative comparisons of different protocols and also provides a useful tool for the optimization of key parameters in whole-body positron emission tomography imaging.

Kernel-based learning from both qualitative and quantitative labels: application to prostate cancer diagnosis based on multiparametric MR imaging.

Building an accurate training database is challenging in supervised classification. For instance, in medical imaging, radiologists often delineate malignant and benign tissues without access to the histological ground truth, leading to uncertain data sets. This paper addresses the pattern classification problem arising when available target data include some uncertainty information. Target data considered here are both qualitative (a class label) or quantitative (an estimation of the posterior probability). In this context, usual discriminative methods, such as the support vector machine (SVM), fail either to learn a robust classifier or to predict accurate probability estimates. We generalize the regular SVM by introducing a new formulation of the learning problem to take into account class labels as well as class probability estimates. This original reformulation into a probabilistic SVM (P-SVM) can be efficiently solved by adapting existing flexible SVM solvers. Furthermore, this framework allows deriving a unique learned prediction function for both decision and posterior probability estimation providing qualitative and quantitative predictions. The method is first tested on synthetic data sets to evaluate its properties as compared with the classical SVM and fuzzy-SVM. It is then evaluated on a clinical data set of multiparametric prostate magnetic resonance images to assess its performances in discriminating benign from malignant tissues. P-SVM is shown to outperform classical SVM as well as the fuzzy-SVM in terms of probability predictions and classification performances, and demonstrates its potential for the design of an efficient computer-aided decision system for prostate cancer diagnosis based on multiparametric magnetic resonance (MR) imaging.

A virtual imaging platform for multi-modality medical image simulation.

This paper presents the Virtual Imaging Platform (VIP), a platform accessible at http://vip.creatis.insa-lyon.fr to facilitate the sharing of object models and medical image simulators, and to provide access to distributed computing and storage resources. A complete overview is presented, describing the ontologies designed to share models in a common repository, the workflow template used to integrate simulators, and the tools and strategies used to exploit computing and storage resources. Simulation results obtained in four image modalities and with different models show that VIP is versatile and robust enough to support large simulations. The platform currently has 200 registered users who consumed 33 years of CPU time in 2011.

Computer-aided diagnosis of prostate cancer in the peripheral zone using multiparametric MRI.

This study evaluated a computer-assisted diagnosis (CADx) system for determining a likelihood measure of prostate cancer presence in the peripheral zone (PZ) based on multiparametric magnetic resonance (MR) imaging, including T2-weighted, diffusion-weighted and dynamic contrast-enhanced MRI at 1.5 T. Based on a feature set derived from grey-level images, including first-order statistics, Haralick features, gradient features, semi-quantitative and quantitative (pharmacokinetic modelling) dynamic parameters, four kinds of classifiers were trained and compared: nonlinear support vector machine (SVM), linear discriminant analysis, k-nearest neighbours and naïve Bayes classifiers. A set of feature selection methods based on t-test, mutual information and minimum-redundancy-maximum-relevancy criteria were also compared. The aim was to discriminate between the relevant features as well as to create an efficient classifier using these features. The diagnostic performances of these different CADx schemes were evaluated based on a receiver operating characteristic (ROC) curve analysis. The evaluation database consisted of 30 sets of multiparametric MR images acquired from radical prostatectomy patients. Using histologic sections as the gold standard, both cancer and nonmalignant (but suspicious) tissues were annotated in consensus on all MR images by two radiologists, a histopathologist and a researcher. Benign tissue regions of interest (ROIs) were also delineated in the remaining prostate PZ. This resulted in a series of 42 cancer ROIs, 49 benign but suspicious ROIs and 124 nonsuspicious benign ROIs. From the outputs of all evaluated feature selection methods on the test bench, a restrictive set of about 15 highly informative features coming from all MR sequences was discriminated, thus confirming the validity of the multiparametric approach. Quantitative evaluation of the diagnostic performance yielded a maximal area under the ROC curve (AUC) of 0.89 (0.81-0.94) for the discrimination of the malignant versus nonmalignant tissues and 0.82 (0.73-0.90) for the discrimination of the malignant versus suspicious tissues when combining the t-test feature selection approach with a SVM classifier. A preliminary comparison showed that the optimal CADx scheme mimicked, in terms of AUC, the human experts in differentiating malignant from suspicious tissues, thus demonstrating its potential for assisting cancer identification in the PZ.

Magnetic resonance imaging follow-up of liver growth of neuroendocrine tumors in an experimental mouse model.

Liver metastases in patients with gastroenteropancreatic (GEP) endocrine tumors represent the main factor of adverse prognosis in this tumor type and thus have a strong effect on the therapeutic strategies. Currently, magnetic resonance imaging (MRI) is considered the modality of choice for the noninvasive, in vivo detection of liver metastases. Dedicated MRI protocols suitable for following liver lesion evolution on an experimental model of endocrine tumors could be valuable. An experimental animal model mimicking the clinical situation of intrahepatic dissemination has been designed. The goal of this study was to characterize liver lesions in this athymic nude mouse model and assess the detection sensitivity of MRI using a physiological gating strategy optimized for high magnetic fields. The experiments were performed at 7 T using a dual cardiac-respiratory-triggered multiple spin-echo sequence. This protocol was used to carry out a longitudinal follow-up of hepatic lesions in a group of eight nude mice at different stages: Day 7 (D7), Day 12 (D12), Day 17 (D17) and Day 24 (D24). The hepatic lesion volume fraction (HLVF) was quantified using an adaptive segmentation procedure based on a dual-reference limit. Mean transverse relaxation time T(2) values were quantified from multiple spin-echo images. The first lesions were detected at stage D12 on images with 20-ms TE. From D12, the HLVF increased significantly with stage. The mean T(2) values also increased significantly at D17 and D24. In conclusion, the level of detection and characterization of liver lesions were performed using a devoted protocol with a dedicated high-field MRI synchronization strategy. In future studies, MRI could be used to monitor the effects of targeted therapies on liver endocrine metastases in preclinical animal models.