Abstract Neoadjuvant chemotherapy is a standard clinical practice for tumor downsizing in breast cancer, with $$^{18}$$ F-FDG Positron Emission Tomography (PET) being an essential tool for predicting complete pathological response and monitoring treatment. Our work aims to leverage PET imaging for automated segmentation of breast lesions and biomarker analysis before and after the first course of chemotherapy. We developed a system to segment primary tumor regions and extract biomarkers reflecting tumor evolution. A total of 243 baseline and 180 follow-up $$^{18}$$ F-FDG PET scans were acquired. Ground truth annotations were generated semi-automatically for all baseline scans and manually for 12 test cases. A deep learning-based segmentation method was developed and evaluated using various architectures. The optimal baseline model, trained on baseline exams, was fine-tuned on 15 follow-up scans using active learning to segment tumors in follow-up exams. The pipeline extracts maximum standardized uptake value (SUV $$_{max}$$ ), metabolic tumor volume (MTV), and total lesion glycolysis (TLG) to assess tumor response. Quality control procedures were used to exclude outliers. Among the tested approaches, nnUNet achieved the best tumor segmentation on PET baseline scans, with a Dice similarity coefficient of $$0.89 \pm 0.04$$ and a Hausdorff distance of $$3.52 \pm 0.76$$ mm on the test set. After fine-tuning, performance on follow-up exams reached a Dice similarity coefficient of $$0.78 \pm 0.03$$ and Hausdorff distance of $$4.95 \pm 0.12$$ mm. Biomarker analysis showed strong correlations between manual and automatic segmentations. The average $$\Delta SUV_{max}$$ , $$\Delta MTV$$ , and $$\Delta TLG$$ between 12 baseline and follow-up scans used for testing were $$-6.02 \pm 1.55$$ (p=0.002), $$-9.30 \pm 2.33$$ cm $$^3$$ (p=0.007), and $$-14.09 \pm 6.33$$ cm $$^3$$ (p=0.010), respectively. The proposed method provides an effective automated system for breast tumor segmentation from $$^{18}$$ F-FDG PET. Thanks to biomarkers extracted from the automatic segmentations on both baseline and follow-up exams, our method enables the automatic assessment of cancer progression.

The aim of this study was to optimise the acquisition and reconstruction parameters for ^90Y imaging on a digital PET scanner (Discovery MI 4-ring) under high (3 GBq), intermediate (1 GBq), and low (200 MBq) activity conditions. First, quantitative linearity was evaluated. Then, NEMA IEC body phantom acquisitions were reconstructed using various scan durations and Q.Clear reconstruction parameters. Next, optimal protocols were identified by minimising the discrepancies between absorbed dose maps derived from images (using the Local Deposition Model (LDM)) and those obtained from Monte Carlo (MC) simulations. The images reconstructed with these optimal protocols were then used to evaluate effective spatial resolution and to compare the accuracy of LDM with that of the Dose Voxel Kernel (DVK) approach. Quantitative linearity analysis revealed an underestimation of phantom activity at high activity (up to − 15.7

Neoadjuvant chemotherapy is a standard clinical practice for tumor downsizing in breast cancer, with [Formula: see text]F-FDG Positron Emission Tomography (PET) being an essential tool for predicting complete pathological response and monitoring treatment. Our work aims to leverage PET imaging for automated segmentation of breast lesions and biomarker analysis before and after the first course of chemotherapy. We developed a system to segment primary tumor regions and extract biomarkers reflecting tumor evolution. A total of 243 baseline and 180 follow-up [Formula: see text]F-FDG PET scans were acquired. Ground truth annotations were generated semi-automatically for all baseline scans and manually for 12 test cases. A deep learning-based segmentation method was developed and evaluated using various architectures. The optimal baseline model, trained on baseline exams, was fine-tuned on 15 follow-up scans using active learning to segment tumors in follow-up exams. The pipeline extracts maximum standardized uptake value (SUV[Formula: see text]), metabolic tumor volume (MTV), and total lesion glycolysis (TLG) to assess tumor response. Quality control procedures were used to exclude outliers. Among the tested approaches, nnUNet achieved the best tumor segmentation on PET baseline scans, with a Dice similarity coefficient of [Formula: see text] and a Hausdorff distance of [Formula: see text] mm on the test set. After fine-tuning, performance on follow-up exams reached a Dice similarity coefficient of [Formula: see text] and Hausdorff distance of [Formula: see text] mm. Biomarker analysis showed strong correlations between manual and automatic segmentations. The average [Formula: see text], [Formula: see text], and [Formula: see text] between 12 baseline and follow-up scans used for testing were [Formula: see text] (p=0.002), [Formula: see text] cm[Formula: see text] (p=0.007), and [Formula: see text] cm[Formula: see text] (p=0.010), respectively. The proposed method provides an effective automated system for breast tumor segmentation from [Formula: see text]F-FDG PET. Thanks to biomarkers extracted from the automatic segmentations on both baseline and follow-up exams, our method enables the automatic assessment of cancer progression.

Selective internal radiation therapy (SIRT) increasingly relies on accurate magnetic resonance imaging (MRI) to computed tomography (CT) registration and accurate liver and tumour segmentation, for effective pre-treatment planning. This study evaluates the impact of automatic registration and segmentation techniques on dosimetry calculations for SIRT treatment planning. It compares semi-automatic and automatic registration, as well as manual and automatic deep learning-based segmentation. Pre-treatment data from 90 patients with hepatocellular carcinoma were analysed. The dataset consisted of contrast-enhanced T1-weighted MRI scans with manually delineated liver and tumour volumes of interest (VOIs), as well as single-photon emission computed tomography (SPECT)/CT scans with manually or semi-automatically delineated liver VOI. The clinical routine pipeline, which involves semi-automatic registration and manual/semi-automatic segmentation, was used as the baseline pipeline and compared to experimental pipelines that use intensity-based deformable automatic registration or deep learning-based automatic segmentation. Dosimetric accuracy was assessed via metrics such as the mean absorbed dose, the minimum dose received by 70

Given the increasing data requirements of deep learning models and the scarcity of medical imaging data, new data augmentation techniques are receiving particular attention. This paper explores the subfield of tumour synthesis within medical image generation, focusing on the development of synthetic tumours in MR images. This study introduces a novel tumour generation method using diffusion models, designed to inpaint visually convincing 3D synthetic liver tumours into real MRI volumes while generating the corresponding masks using simplex deformation. This approach has been employed successfully to inpaint images with 1000 synthetic tumours. Furthermore, it has shown significant performance improvements when applied in image segmentation tasks. In particular, our method improved the Dice coefficient by 6.7 points on the ATLAS test set without relying on external data. When combined with a pseudo-annotated external dataset, the improvement increased to 10 points. This study not only demonstrates the ability to segment tumours but also paves the way for various synthetic data-based applications in medical imaging.

This study aims to investigate whether quantitative radiomic features extracted from Positron Emission Tomography/Computed Tomography (PET/CT) could differentiate triple-negative breast cancer (TNBC) from non-triple-negative breast cancer (non-TNBC). We propose a pipeline that combines deep learning for cancer lesion segmentation with machine learning techniques to classify TNBC. Our approach leveraged the radiomic features extracted from 18F-fluorodeoxyglucose PET/CT. This retrospective study included the PET/CT images of 217 patients with breast cancer (57 TNBC and 160 non-TNBC) admitted to Georges-François Leclerc Hospital. The tumor regions of interest were automatically segmented on PET images using a deep learning model and mapped to CT scans. Radiomic features were extracted from 3D tumor volumes and machine learning classifiers were built using stratified 5-fold cross-validation. Recursive feature elimination was employed to rank and select the most relevant radiomic features, thereby enhancing classification performance. The model was evaluated using the F1-score, area under the receiver operating characteristic (ROC) curve (AUC), accuracy, sensitivity, and specificity. The proposed method achieved promising performance, with an F1-score of 0.90 ± 0.02, an accuracy of 0.86 ± 0.07, a sensitivity of 0.91 ± 0.06, and an AUC of 0.88 ± 0.04, using the top-ranked features. The metrics were evaluated as the average over a five-fold cross-validation. Radiomic features extracted from PET and CT scans provide valuable prognostic insights for the identification of TNBC. This study demonstrated that machine learning algorithms based on radiomic features and automated PET/CT segmentation can accurately distinguish TNBC from non-TNBC.Clinical relevance— This study demonstrates the potential of image-based radiomic analysis combined with machine learning to differentiate triple-negative breast cancer (TNBC) from non-TNBC. By using deep learning for automatic tumor segmentation and feature extraction, this approach offers a non-invasive, quantitative tool that can improve TNBC diagnosis and the efficiency of treatment strategies. These advancements may help clinicians provide more reliable insights, while reducing the likelihood of misclassification.

Neoadjuvant chemotherapy (NAC) has become a standard clinical practice for tumor downsizing in breast cancer with 18F-FDG Positron Emission Tomography (PET). Our work aims to leverage PET imaging for the segmentation of breast lesions. The focus is on developing an automated system that accurately segments primary tumor regions and extracts key biomarkers from these areas to provide insights into the evolution of breast cancer following the first course of NAC. 243 baseline 18F-FDG PET scans (PET_Bl) and 180 follow-up 18F-FDG PET scans (PET_Fu) were acquired before and after the first course of NAC, respectively. Firstly, a deep learning-based breast tumor segmentation method was developed. The optimal baseline model (model trained on baseline exams) was fine-tuned on 15 follow-up exams and adapted using active learning to segment tumor areas in PET_Fu. The pipeline computes biomarkers such as maximum standardized uptake value (SUVmax), metabolic tumor volume (MTV), and total lesion glycolysis (TLG) to evaluate tumor evolution between PET_Fu and PET_Bl. Quality control measures were employed to exclude aberrant outliers. The nnUNet deep learning model outperformed in tumor segmentation on PET_Bl, achieved a Dice similarity coefficient (DSC) of 0.89 and a Hausdorff distance (HD) of 3.52 mm. After fine-tuning, the model demonstrated a DSC of 0.78 and a HD of 4.95 mm on PET_Fu exams. Biomarkers analysis revealed very strong correlations whatever the biomarker between manually segmented and automatically predicted regions. The significant average decrease of SUVmax, MTV and TLG were 5.22, 11.79 cm3 and 19.23 cm3, respectively. The presented approach demonstrates an automated system for breast tumor segmentation from 18F-FDG PET. Thanks to the extracted biomarkers, our method enables the automatic assessment of cancer progression.

This study aimed to assess the potential of 18FFluorodeoxyglucose Positron Emission Tomography/Computed Tomography (18F-FDG PET/CT) parameters, including advanced texture features, to predict pathological complete response (pCR) after the first course of neoadjuvant chemotherapy (NAC) in breast cancer follow-up patients. This approach evaluated pCR after the first course of NAC by combining information from functional images, anatomical images, and clinical data. A total of 204 breast cancer patients underwent 18F-FDG PET/CT imaging for NAC assessment. From these delayed PET/CT scans, we extracted both metabolic and radiomic features, combining imaging parameters with the breast cancer molecular subtype information for each patient to improve pCR prediction. Lesion segmentation was automated using the no-new-Net (nnUNet) deep learning model. To predict pCR, we employed machine learning classifiers, including Random Forest, XGBoost, and Support Vector Machine. Among all tested models, the highest prediction performance was achieved when PET/CT features (both baseline and follow-up) were combined with breast cancer subtype information. The analysis was conducted on the entire dataset (Human Epidermal Growth Factor Receptor 2 (HER2), Luminal, and Triple-Negative (TN). Moreover, separate analyses were performed specifically on HER2 tumors (N=76) and TN tumors (N=52). The combined model achieved a mean balanced accuracy of 0.76 ± 0.09, surpassing the individual models for HER2 (0.67 ± 0.08) and TN (0.65 ± 0.06). These findings show the importance of integrating baseline and follow-up PET/CT radiomic features, texture analysis, and clinical information for more accurate pCR prediction after the first course of NAC in breast cancer patients. Overall, the features extracted from baseline data and follow-up data or after the first course of NAC, combined with information of breast cancer subtype, offer strong predictive value for pCR in follow-up patients.Clinical Relevance—By providing a more accurate assessment of treatment response after the first course of NAC, this approach empowers clinicians to make artificial intelligence-driven decisions, customize therapy plans for individual patients, and avoid ineffective treatments. Consequently, this strategy could improve patient outcomes and optimize therapeutic efficacy.

In medical imaging, accurate segmentation is crucial to improving diagnosis, treatment, or both. However, navigating the multitude of available architectures for automatic segmentation can be overwhelming, making it challengingto determine the appropriate type of architecture and tune the most crucial parameters during dataset optimisation.To address this problem, we examined and refined seven distinct architectures for segmenting the liver, as well as liver tumours, with a restricted training collection of 60 3D contrast-enhanced magnetic resonance images (CE-MRI) from the ATLAS dataset. Included in these architectures are convolutional neural networks (CNNs), transformers, and hybrid CNN/transformer architectures. Bayesian search techniques were used for hyperparameter tuning to hasten convergence to the optimal parameter mixes while also minimising the number of trained models. It was unexpected that hybrid models, which typically exhibit superior performance on larger datasets, would exhibit comparable performance to CNNs. The optimisation of parameters contributed to better segmentations, resulting in an average increase of 1.7% and 5.0% in liver and tumour segmentation Dice coefficients, respectively.In conclusion, the findings of this study indicate that hybrid CNN/transformer architectures may serve as a practical substitute for CNNs even in small datasets. This underscores the significance of hyperparameter optimisation.

Background We propose a comprehensive evaluation of a Discovery MI 4-ring (DMI) model, using a Monte Carlo simulator (GATE) and a clinical reconstruction software package (PET toolbox). The following performance characteristics were compared with actual measurements according to NEMA NU 2-2018 guidelines: system sensitivity, count losses and scatter fraction (SF), coincidence time resolution (CTR), spatial resolution (SR), and image quality (IQ). For SR and IQ tests, reconstruction of time-of-flight (TOF) simulated data was performed using the manufacturer’s reconstruction software. Results Simulated prompt, random, true, scatter and noise equivalent count rates closely matched the experimental rates with maximum relative differences of 1.6%, 5.3%, 7.8%, 6.6%, and 16.5%, respectively, in a clinical range of less than 10 kBq/mL. A 3.6% maximum relative difference was found between experimental and simulated sensitivities. The simulated spatial resolution was better than the experimental one. Simulated image quality metrics were relatively close to the experimental results. Conclusions The current model is able to reproduce the behaviour of the DMI count rates in the clinical range and generate clinical-like images with a reasonable match in terms of contrast and noise.

Objective. We introduce a versatile methodology for the accurate modelling of PET imaging systems via Monte Carlo simulations, using the Geant4 application for tomographic emission (GATE) platform. Accurate Monte Carlo modelling involves the incorporation of a complete analytical signal processing chain, called the digitizer in GATE, to emulate the different count rates encountered in actual positron emission tomography (PET) systems.Approach. The proposed approach consists of two steps: (1) modelling the digitizer to replicate the detection chain of real systems, covering all available parameters, whether publicly accessible or supplied by manufacturers; (2) estimating the remaining parameters, i.e. background noise level, detection efficiency, and pile-up, using optimisation techniques based on experimental single and prompt event rates. We show that this two-step optimisation reproduces the other experimental count rates (true, scatter, and random), without the need for additional adjustments. This method has been applied and validated with experimental data derived from the NEMA count losses test for three state-of-the-art SiPM-based time-of-flight (TOF)-PET systems: Philips Vereos, Siemens Biograph Vision 600 and GE Discovery MI 4-ring.Main results. The results show good agreement between experiments and simulations for the three PET systems, with absolute relative discrepancies below 3%, 6%, 6%, 7% and 12% for prompt, random, true, scatter and noise equivalent count rates, respectively, within the 0-10 kBq·ml-1activity concentration range typically observed in whole-body18F scans.Significance. Overall, the proposed digitizer optimisation method was shown to be effective in reproducing count rates and NECR for three of the latest generation SiPM-based TOF-PET imaging systems. The proposed methodology could be applied to other PET scanners.

Selective internal radiation therapy (SIRT) is a targeted treatment for liver tumors, particularly hepatocellular carcinoma (HCC), that involves the precise delivery of radioactive microspheres to the tumor's blood supply. Image registration is critical to SIRT, ensuring precise alignment of diagnostic images with the treatment plan to optimize microsphere delivery and minimize side effects. In the context of SIRT, our goal was to develop a fully-automatic hybrid registration pipeline, using liver segmentation masks from a 3D UNet model to achieve performance comparable to expert registrations. The pipeline combines conventional and deep learning methods for automatic alignment of pre-treatment magnetic resonance images (MRI) on single-photon emission computed tomography (SPECT)/computed tomography (CT) images. This hybrid pipeline, which uses conventional global rigid registration and a deep learning-based approach for local deformation, outperformed conventional and manual expert registration. Quantitative assessment on a dataset of 69 HCC patients showed an improved Dice similarity coefficient (DSC) of 0.928 compared to 0.917 with the conventional methods. A subset analysis of 61 patients with expert registrations showed a mean DSC of 0.922, while our proposed method remained at a mean DSC of 0.928. These results demonstrate the effectiveness of our hybrid approach in achieving accurate liver registration, which is critical for precise microsphere delivery during SIRT. The improvement over conventional methods highlights the potential of incorporating deep learning techniques into multimodal liver registration, thereby improving the overall quality and effectiveness of SIRT in the management of HCC. Our method provides clinicians with a reliable and automated registration pipeline that can positively optimize treatment planning and reduce the burden of manual registration. As a result, our hybrid approach holds promise for more accurate and precise registration results.

BACKGROUND:The aim of this study is to investigate the added value of combining tumour blood flow (BF) and metabolism parameters, including texture features, with clinical parameters to predict, at baseline, the pathological complete response (pCR) to neoadjuvant chemotherapy (NAC) in patients with newly diagnosed breast cancer (BC). METHODS:One hundred and twenty-eight BC patients underwent a 18F-FDG PET/CT before any treatment. Tumour BF and metabolism parameters were extracted from first-pass dynamic and delayed PET images, respectively. Standard and texture features were extracted from BF and metabolic images. Prediction of pCR was performed using logistic regression, random forest and support vector classification algorithms. Models were built using clinical (C), clinical and metabolic (C+M) and clinical, metabolic and tumour BF (C+M+BF) information combined. Algorithms were trained on 80% of the dataset and tested on the remaining 20%. Univariate and multivariate features selections were carried out on the training dataset. A total of 50 shuffle splits were performed. The analysis was carried out on the whole dataset (HER2 and Triple Negative (TN)), and separately in HER2 (N=76) and TN (N=52) tumours. RESULTS:In the whole dataset, the highest classification performances were observed for C+M models, significantly (p-value<0.01) higher than C models and better than C+M+BF models (mean balanced accuracy of 0.66, 0.61, and 0.64 respectively). For HER2 tumours, equal performances were noted for C and C+M models, with performances higher than C+M+BF models (mean balanced accuracy of 0.64, and 0.61 respectively). Regarding TN tumours, the best classification results were reported for C+M models, with better performances than C and C+M+BF models but not significantly (mean balanced accuracy of 0.65, 0.63, and 0.62 respectively). CONCLUSION:Baseline clinical data combined with global and texture tumour metabolism parameters assessed by 18F-FDG PET/CT provide a better prediction of pCR after NAC in patients with BC compared to clinical parameters alone for TN, and HER2 and TN tumours together. In contrast, adding BF parameters to the models did not improve prediction, regardless of the tumour subgroup analysed.

With the development of three-dimensional (3D) printing, 3D-printed products have been widely used in medical fields, such as plastic surgery, orthopedics, dentistry, etc. In cardiovascular research, 3D-printed models are becoming more realistic in shape. However, from a biomechanical point of view, only a few studies have explored printable materials that can represent the properties of the human aorta. This study focuses on 3D-printed materials that might simulate the stiffness of human aortic tissue. First, the biomechanical properties of a healthy human aorta were defined and used as reference. The main objective of this study was to identify 3D printable materials that possess similar properties to the human aorta. Three synthetic materials, NinjaFlex (Fenner Inc., Manheim, USA), FilasticTM (Filastic Inc., Jardim Paulistano, Brazil), and RGD450+TangoPlus (Stratasys Ltd.©, Rehovot, Israel), were printed in different thicknesses. Uniaxial and biaxial tensile tests were performed to compute several biomechanical properties, such as thickness, stress, strain, and stiffness. We found that with the mixed material RGD450+TangoPlus, it was possible to achieve a similar stiffness to healthy human aorta. Moreover, the 50-shore-hardness RGD450+TangoPlus had similar thickness and stiffness to the human aorta.

Purpose: The aim of this study was (a) to optimise the 99mTc-SPECT reconstruction parameters for the pre-treatment dosimetry of 90Y-selective internal radiation therapy (SIRT) and (b) to compare the accuracy of clinical dosimetry methods with full Monte-Carlo dosimetry (fMCD) performed with Gate.Methods: To optimise the reconstruction parameters, two hundred reconstructions with different parameters were performed on a NEMA phantom, varying the number of iterations, subsets, and post-filtering. The accuracy of the dosimetric methods was then investigated using an anthropomorphic phantom. Absorbed dose maps were generated using (1) the Partition Model (PM), (2) the Dose Voxel Kernel (DVK) convolution, and (3) the Local Deposition Method (LDM) with known activity restricted to the whole phantom (WP) or to the liver and lungs (LL). The dose to the lungs was calculated using the "multiple DVK"and "multiple LDM"methods.Results: Optimal OSEM reconstruction parameters were found to depend on object size and dosimetric criterion chosen (Dmean or DVH-derived metric). The Dmean of all three dosimetric methods was close (<= 10%) to the Dmean of fMCD simulations when considering large segmented volumes (whole liver, normal liver). In contrast, the Dmean to the small volume ( null =31) was systemically underestimated (12%-25%). For lungs, the "multiple DVK"and "multiple LDM"methods yielded a Dmean within 20% for the WP method and within 10% for the LL method.Conclusions: All three methods showed a substantial degradation of the dose-volume histograms (DVHs) compared to fMCD simulations. The DVK and LDM methods performed almost equally well, with the "multiple DVK"method being more accurate in the lungs.

Ascending aortic aneurysm is a pathology that is important to be supervised and treated. During the years the aorta dilates, it becomes stiff, and its elastic properties decrease. In some cases, the aortic wall can rupture leading to aortic dissection with a high mortality rate. The main reference standard to measure when the patient needs to undertake surgery is the aortic diameter. However, the aortic diameter was shown not to be sufficient to predict aortic dissection, implying other characteristics should be considered. Therefore, the main objective of this work is to assess in-vivo the elastic properties of four different quadrants of the ascending aorta and compare the results with equivalent properties obtained ex-vivo. The database consists of 73 cine-MRI sequences of thoracic aorta acquired in axial orientation at the level of the pulmonary trunk. All the patients have dilated aorta and surgery is required. The exams were acquired just prior to surgery, each consisting of 30 slices on average across the cardiac cycle. Multiple deep learning architectures have been explored with different hyperparameters and settings to automatically segment the contour of the aorta on each image and then automatically calculate the aortic compliance. A semantic segmentation U-Net network outperforms the rest explored networks with a Dice score of 98.09% (±0.96%) and a Hausdorff distance of 4.88 mm (±1.70 mm). Local aortic compliance and local aortic wall strain were calculated from the segmented surfaces for each quadrant and then compared with elastic properties obtained ex-vivo. Good agreement was observed between Young's modulus and in-vivo strain. Our results suggest that the lateral and posterior quadrants are the stiffest. In contrast, the medial and anterior quadrants have the lowest aortic stiffness. The in-vivo stiffness tendency agrees with the values obtained ex-vivo. We can conclude that our automatic segmentation method is robust and compatible with clinical practice (thanks to a graphical user interface), while the in-vivo elastic properties are reliable and compatible with the ex-vivo ones.

A thoracic aortic aneurysm is an abnormal dilatation of the aorta that can progress and lead to rupture. The decision to conduct surgery is made by considering the maximum diameter, but it is now well known that this metric alone is not completely reliable. The advent of 4D flow magnetic resonance imaging has allowed for the calculation of new biomarkers for the study of aortic diseases, such as wall shear stress. However, the calculation of these biomarkers requires the precise segmentation of the aorta during all phases of the cardiac cycle. The objective of this work was to compare two different methods for automatically segmenting the thoracic aorta in the systolic phase using 4D flow MRI. The first method is based on a level set framework and uses the velocity field in addition to 3D phase contrast magnetic resonance imaging. The second method is a U-Net-like approach that is only applied to magnitude images from 4D flow MRI. The used dataset was composed of 36 exams from different patients, with ground truth data for the systolic phase of the cardiac cycle. The comparison was performed based on selected metrics, such as the Dice similarity coefficient (DSC) and Hausdorf distance (HD), for the whole aorta and also three aortic regions. Wall shear stress was also assessed and the maximum wall shear stress values were used for comparison. The U-Net-based approach provided statistically better results for the 3D segmentation of the aorta, with a DSC of 0.92 ± 0.02 vs. 0.86 ± 0.5 and an HD of 21.49 ± 24.8 mm vs. 35.79 ± 31.33 mm for the whole aorta. The absolute difference between the wall shear stress and ground truth slightly favored the level set method, but not significantly (0.754 ± 1.07 Pa vs. 0.737 ± 0.79 Pa). The results showed that the deep learning-based method should be considered for the segmentation of all time steps in order to evaluate biomarkers based on 4D flow MRI.

Liver cancer is the sixth most common cancer in the world and the fourth leading cause of cancer mortality. In unresectable liver cancers, especially hepatocellular carcinoma (HCC), transarterial radioembolisation (TARE) can be considered for treatment. TARE treatment involves a contrast-enhanced magnetic resonance imaging (CE-MRI) exam performed beforehand to delineate the liver and tumour(s) in order to perform dosimetry calculation. Due to the significant amount of time and expertise required to perform the delineation process, there is a strong need for automation. Unfortunately, the lack of publicly available CE-MRI datasets with liver tumour annotations has hindered the development of fully automatic solutions for liver and tumour segmentation. The “Tumour and Liver Automatic Segmentation” (ATLAS) dataset that we present consists of 90 liver-focused CE-MRI covering the entire liver of 90 patients with unresectable HCC, along with 90 liver and liver tumour segmentation masks. To the best of our knowledge, the ATLAS dataset is the first public dataset providing CE-MRI of HCC with annotations. The public availability of this dataset should greatly facilitate the development of automated tools designed to optimise the delineation process, which is essential for treatment planning in liver cancer patients.

BACKGROUND:4D flow MRI allows the analysis of hemodynamic changes in the aorta caused by pathologies such as thoracic aortic aneurysms (TAA). For personalized management of TAA, new biomarkers are required to analyze the effect of fluid structure iteration which can be obtained from 4D flow MRI. However, the generation of these biomarkers requires prior 4D segmentation of the aorta.OBJECTIVE:To develop an automatic deep learning model to segment the aorta in 4D from 4D flow MRI.METHODS:Segmentation is addressed with a U-Net based segmentation model that treats each 4D flow MRI frame as an independent sample. Performance is measured with respect to Dice score (DS) and Hausdorff distance (HD). In addition, the maximum and minimum surface areas at the level of the ascending aorta are measured and compared with those obtained from cine-MRI.RESULTS:The segmentation performance was 0.90 ± 0.02 for the DS and the mean HD was 9.58 ± 4.36 mm. A correlation coefficient of r = 0.85 was obtained for the maximum surface and r = 0.86 for the minimum surface between the 4D flow MRI and cine-MRI.CONCLUSION:The proposed automatic approach of 4D aortic segmentation from 4D flow MRI seems to be accurate enough to contribute to the wider use of this imaging technique in the analysis of pathologies such as TAA.

In the management of the aortic aneurysm, 4D flow magnetic resonance Imaging provides valuable information for the computation of new biomarkers using computational fluid dynamics (CFD). However, accurate segmentation of the aorta is required. Thus, our objective is to evaluate the performance of two automatic segmentation methods on the calculation of aortic wall pressure. Automatic segmentation of the aorta was performed with methods based on deep learning and multi-atlas using the systolic phase in the 4D flow MRI magnitude image of 36 patients. Using mesh morphing, isotopological meshes were generated, and CFD was performed to calculate the aortic wall pressure. Node-to-node comparisons of the pressure results were made to identify the most robust automatic method respect to the pressures obtained with a manually segmented model. Deep learning approach presented the best segmentation performance with a mean Dice similarity coefficient and a mean Hausdorff distance (HD) equal to 0.92+/− 0.02 and 21.02+/− 24.20 mm, respectively. At the global level HD is affected by the performance in the abdominal aorta. Locally, this distance decreases to 9.41+/− 3.45 and 5.82+/− 6.23 for the ascending and descending thoracic aorta, respectively. Moreover, with respect to the pressures from the manual segmentations, the differences in the pressures computed from deep learning were lower than those computed from multi-atlas method. To reduce biases in the calculation of aortic wall pressure, accurate segmentation is needed, particularly in regions with high blood flow velocities. Thus, the deep learning segmen-tation method should be preferred.