Cooperative AI training for cardiothoracic segmentation in computed tomography: An iterative multi-center annotation approach.

PURPOSE: Radiological reporting is transitioning to quantitative analysis, requiring large-scale multi-center validation of biomarkers. A major prerequisite and bottleneck for this task is the voxelwise annotation of image data, which is time-consuming for large cohorts. In this study, we propose an iterative training workflow to support and facilitate such segmentation tasks, specifically for high-resolution thoracic CT data. METHODS: Our study included 132 thoracic CT scans from clinical practice, annotated by 13 radiologists. In three iterative training experiments, we aimed to improve and accelerate segmentation of the heart and mediastinum. Each experiment started with manual segmentation of 5-25 CT scans, which served as training data for a nnU-Net. Further iterations incorporated AI pre-segmentation and human correction to improve accuracy, accelerate the annotation process, and reduce human involvement over time. RESULTS: Results showed consistent improvement in AI model quality with each iteration. Resampled datasets improved the Dice similarity coefficients for both the heart (DCS 0.91 [0.88; 0.92]) and the mediastinum (DCS 0.95 [0.94; 0.95]). Our AI models reduced human interaction time by 50 % for heart and 70 % for mediastinum segmentation in the most potent iteration. A model trained on only five datasets achieved satisfactory results (DCS > 0.90). CONCLUSIONS: The iterative training workflow provides an efficient method for training AI-based segmentation models in multi-center studies, improving accuracy over time and simultaneously reducing human intervention. Future work will explore the use of fewer initial datasets and additional pre-processing methods to enhance model quality.

Deep learning-based segmentation of brain parenchyma and ventricular system in CT scans in the presence of anomalies.

Introduction: The automatic segmentation of brain parenchyma and cerebrospinal fluid-filled spaces such as the ventricular system is the first step for quantitative and qualitative analysis of brain CT data. For clinical practice and especially for diagnostics, it is crucial that such a method is robust to anatomical variability and pathological changes such as (hemorrhagic or neoplastic) lesions and chronic defects. This study investigates the increase in overall robustness of a deep learning algorithm that is gained by adding hemorrhage training data to an otherwise normal training cohort. Methods: A 2D U-Net is trained on subjects with normal appearing brain anatomy. In a second experiment the training data includes additional subjects with brain hemorrhage on image data of the RSNA Brain CT Hemorrhage Challenge with custom reference segmentations. The resulting networks are evaluated on normal and hemorrhage test casesseparately, and on an independent test set of patients with brain tumors of the publicly available GLIS-RT dataset. Results: Adding data with hemorrhage to the training set significantly improves the segmentation performance over an algorithm trained exclusively on normally appearing data, not only in the hemorrhage test set but also in the tumor test set. The performance on normally appearing data is stable. Overall, the improved algorithm achieves median Dice scores of 0.98 (parenchyma), 0.91 (left ventricle), 0.90 (right ventricle), 0.81 (third ventricle), and 0.80 (fourth ventricle) on the hemorrhage test set. On the tumor test set, the median Dice scores are 0.96 (parenchyma), 0.90 (left ventricle), 0.90 (right ventricle), 0.75 (third ventricle), and 0.73 (fourth ventricle). Conclusion: Training on an extended data set that includes pathologies is crucial and significantly increases the overall robustness of a segmentation algorithm for brain parenchyma and ventricular system in CT data, also for anomalies completely unseen during training. Extension of the training set to include other diseases may further improve the generalizability of the algorithm.

Quantitative evaluation of MR perfusion imaging using blood pool contrast agent in subjects without pulmonary diseases and in patients with pulmonary embolism.

OBJECTIVE: To assess the feasibility of time-resolved parallel three-dimensional magnetic resonance imaging (MRI) for quantitative analysis of pulmonary perfusion using a blood pool contrast agent. METHODS: Quantitative perfusion analysis was performed using novel software to assess pulmonary blood flow (PBF), pulmonary blood volume (PBV) and mean transit time (MTT) in a quantitative manner. RESULTS: The evaluation of lung perfusion in the normal subjects showed an increase of PBF, PBV ventrally to dorsally (gravitational direction), and the highest values at the upper lobe, with a decrease to the middle and lower lobe (isogravitational direction). MTT showed no relevant changes in either the gravitational or isogravitational directions. In comparison with normally perfused lung areas (in diseased patients), the pulmonary embolism (PE) regions showed a significantly lower mean PBF (20 ± 0.6 ml/100 ml/min, normal region 94 ± 1 ml/100 ml/min; P < 0.001), mean PBV (2 ± 0.1 ml/100 ml, normal region 9.8 ± 0.1 ml/100 ml; P < 0.001) and mean MTT (3.8 ± 0.1 s; normal region 6.3 ± 0.1; P < 0.001). CONCLUSION: Our results demonstrate the feasibility of using time-resolved dynamic contrast-enhanced MRI to determine normal range and regional variation of pulmonary perfusion and perfusion deficits in patients with PE. KEY POINTS: • Recently introduced blood pool contrast agents improve MR evaluation of lung perfusion • Regional differences in lung perfusion indicating a gravitational and isogravitational dependency. • Focal areas of significantly decreased perfusion are detectable in pulmonary embolism.

The International Radiomics Platform - An Initiative of the German and Austrian Radiological Societies - First Application Examples. / Die Internationale Radiomics-Plattform ­ eine Initiative der Deutschen und Österreichischen Röntgengesellschaften ­ Erste Anwendungsbeispiele.

PURPOSE: The DRG-ÖRG IRP (Deutsche Röntgengesellschaft-Österreichische Röntgengesellschaft international radiomics platform) represents a web-/cloud-based radiomics platform based on a public-private partnership. It offers the possibility of data sharing, annotation, validation and certification in the field of artificial intelligence, radiomics analysis, and integrated diagnostics. In a first proof-of-concept study, automated myocardial segmentation and automated myocardial late gadolinum enhancement (LGE) detection using radiomic image features will be evaluated for myocarditis data sets. MATERIALS AND METHODS: The DRG-ÖRP IRP can be used to create quality-assured, structured image data in combination with clinical data and subsequent integrated data analysis and is characterized by the following performance criteria: Possibility of using multicentric networked data, automatically calculated quality parameters, processing of annotation tasks, contour recognition using conventional and artificial intelligence methods and the possibility of targeted integration of algorithms. In a first study, a neural network pre-trained using cardiac CINE data sets was evaluated for segmentation of PSIR data sets. In a second step, radiomic features were applied for segmental detection of LGE of the same data sets, which were provided multicenter via the IRP. RESULTS: First results show the advantages (data transparency, reliability, broad involvement of all members, continuous evolution as well as validation and certification) of this platform-based approach. In the proof-of-concept study, the neural network demonstrated a Dice coefficient of 0.813 compared to the expert's segmentation of the myocardium. In the segment-based myocardial LGE detection, the AUC was 0.73 and 0.79 after exclusion of segments with uncertain annotation.The evaluation and provision of the data takes place at the IRP, taking into account the FAT (fairness, accountability, transparency) and FAIR (findable, accessible, interoperable, reusable) criteria. CONCLUSION: It could be shown that the DRG-ÖRP IRP can be used as a crystallization point for the generation of further individual and joint projects. The execution of quantitative analyses with artificial intelligence methods is greatly facilitated by the platform approach of the DRG-ÖRP IRP, since pre-trained neural networks can be integrated and scientific groups can be networked.In a first proof-of-concept study on automated segmentation of the myocardium and automated myocardial LGE detection, these advantages were successfully applied.Our study shows that with the DRG-ÖRP IRP, strategic goals can be implemented in an interdisciplinary way, that concrete proof-of-concept examples can be demonstrated, and that a large number of individual and joint projects can be realized in a participatory way involving all groups. KEY POINTS: · The DRG-ÖRG IRP is a web/cloud-based radiomics platform based on a public-private partnership.. · The DRG-ÖRG IRP can be used for the creation of quality-assured, structured image data in combination with clinical data and subsequent integrated data analysis.. · First results show the applicability of left ventricular myocardial segmentation using a neural network and segment-based LGE detection using radiomic image features.. · The DRG-ÖRG IRP offers the possibility of integrating pre-trained neural networks and networking of scientific groups.. CITATION FORMAT: · Overhoff D, Kohlmann P, Frydrychowicz A et al. The International Radiomics Platform - An Initiative of the German and Austrian Radiological Societies. Fortschr Röntgenstr 2021; 193: 276 - 287.

LiveSync: deformed viewing spheres for knowledge-based navigation.

Although real-time interactive volume rendering is available even for very large data sets, this visualization method is used quite rarely in the clinical practice. We suspect this is because it is very complicated and time consuming to adjust the parameters to achieve meaningful results. The clinician has to take care of the appropriate viewpoint, zooming, transfer function setup, clipping planes and other parameters. Because of this, most often only 2D slices of the data set are examined. Our work introduces LiveSync, a new concept to synchronize 2D slice views and volumetric views of medical data sets. Through intuitive picking actions on the slice, the users define the anatomical structures they are interested in. The 3D volumetric view is updated automatically with the goal that the users are provided with expressive result images. To achieve this live synchronization we use a minimal set of derived information without the need for segmented data sets or data-specific pre-computations. The components we consider are the picked point, slice view zoom, patient orientation, viewpoint history, local object shape and visibility. We introduce deformed viewing spheres which encode the viewpoint quality for the components. A combination of these deformed viewing spheres is used to estimate a good viewpoint. Our system provides the physician with synchronized views which help to gain deeper insight into the medical data with minimal user interaction.

High-level user interfaces for transfer function design with semantics.

Many sophisticated techniques for the visualization of volumetric data such as medical data have been published. While existing techniques are mature from a technical point of view, managing the complexity of visual parameters is still difficult for non-expert users. To this end, this paper presents new ideas to facilitate the specification of optical properties for direct volume rendering. We introduce an additional level of abstraction for parametric models of transfer functions. The proposed framework allows visualization experts to design high-level transfer function models which can intuitively be used by non-expert users. The results are user interfaces which provide semantic information for specialized visualization problems. The proposed method is based on principal component analysis as well as on concepts borrowed from computer animation.

Automatic lung segmentation method for MRI-based lung perfusion studies of patients with chronic obstructive pulmonary disease.

PURPOSE: A novel fully automatic lung segmentation method for magnetic resonance (MR) images of patients with chronic obstructive pulmonary disease (COPD) is presented. The main goal of this work was to ease the tedious and time-consuming task of manual lung segmentation, which is required for region-based volumetric analysis of four-dimensional MR perfusion studies which goes beyond the analysis of small regions of interest. METHODS: The first step in the automatic algorithm is the segmentation of the lungs in morphological MR images with higher spatial resolution than corresponding perfusion MR images. Subsequently, the segmentation mask of the lungs is transferred to the perfusion images via nonlinear registration. Finally, the masks for left and right lungs are subdivided into a user-defined number of partitions. Fourteen patients with two time points resulting in 28 perfusion data sets were available for the preliminary evaluation of the developed methods. RESULTS: Resulting lung segmentation masks are compared with reference segmentations from experienced chest radiologists, as well as with total lung capacity (TLC) acquired by full-body plethysmography. TLC results were available for thirteen patients. The relevance of the presented method is indicated by an evaluation, which shows high correlation between automatically generated lung masks with corresponding ground-truth estimates. CONCLUSION: The evaluation of the developed methods indicates good accuracy and shows that automatically generated lung masks differ from expert segmentations about as much as segmentations from different experts.