Defending against Reconstruction Attacks through Differentially Private Federated Learning for Classification of Heterogeneous Chest X-ray Data.

Privacy regulations and the physical distribution of heterogeneous data are often primary concerns for the development of deep learning models in a medical context. This paper evaluates the feasibility of differentially private federated learning for chest X-ray classification as a defense against data privacy attacks. To the best of our knowledge, we are the first to directly compare the impact of differentially private training on two different neural network architectures, DenseNet121 and ResNet50. Extending the federated learning environments previously analyzed in terms of privacy, we simulated a heterogeneous and imbalanced federated setting by distributing images from the public CheXpert and Mendeley chest X-ray datasets unevenly among 36 clients. Both non-private baseline models achieved an area under the receiver operating characteristic curve (AUC) of 0.94 on the binary classification task of detecting the presence of a medical finding. We demonstrate that both model architectures are vulnerable to privacy violation by applying image reconstruction attacks to local model updates from individual clients. The attack was particularly successful during later training stages. To mitigate the risk of a privacy breach, we integrated Rényi differential privacy with a Gaussian noise mechanism into local model training. We evaluate model performance and attack vulnerability for privacy budgets ε∈{1,3,6,10}. The DenseNet121 achieved the best utility-privacy trade-off with an AUC of 0.94 for ε=6. Model performance deteriorated slightly for individual clients compared to the non-private baseline. The ResNet50 only reached an AUC of 0.76 in the same privacy setting. Its performance was inferior to that of the DenseNet121 for all considered privacy constraints, suggesting that the DenseNet121 architecture is more robust to differentially private training.

Tubular shape aware data generation for segmentation in medical imaging.

PURPOSE: Chest X-ray is one of the most widespread examinations of the human body. In interventional radiology, its use is frequently associated with the need to visualize various tube-like objects, such as puncture needles, guiding sheaths, wires, and catheters. Detection and precise localization of these tube-like objects in the X-ray images are, therefore, of utmost value, catalyzing the development of accurate target-specific segmentation algorithms. Similar to the other medical imaging tasks, the manual pixel-wise annotation of the tubes is a resource-consuming process. METHODS: In this work, we aim to alleviate the lack of annotated images by using artificial data. Specifically, we present an approach for synthetic generation of the tube-shaped objects, with a generative adversarial network being regularized with a prior-shape constraint. Namely, our model uses Frangi-based regularization to draw synthetic tubes in the predefined fake mask regions and, then, uses the adversarial component to preserve the global realistic appearance of the synthesized image. RESULTS: Our method eliminates the need for the paired image-mask data and requires only a weakly labeled dataset, with fine-tuning on a small paired sample (10-20 images) proving sufficient to reach the accuracy of the fully supervised models. CONCLUSION: We report the applicability of the approach for the task of segmenting tubes and catheters in the X-ray images, whereas the results should also hold for the other acquisition modalities and image computing applications that contain tubular objects.

Smart chest X-ray worklist prioritization using artificial intelligence: a clinical workflow simulation.

OBJECTIVE: The aim is to evaluate whether smart worklist prioritization by artificial intelligence (AI) can optimize the radiology workflow and reduce report turnaround times (RTATs) for critical findings in chest radiographs (CXRs). Furthermore, we investigate a method to counteract the effect of false negative predictions by AI-resulting in an extremely and dangerously long RTAT, as CXRs are sorted to the end of the worklist. METHODS: We developed a simulation framework that models the current workflow at a university hospital by incorporating hospital-specific CXR generation rates and reporting rates and pathology distribution. Using this, we simulated the standard worklist processing "first-in, first-out" (FIFO) and compared it with a worklist prioritization based on urgency. Examination prioritization was performed by the AI, classifying eight different pathological findings ranked in descending order of urgency: pneumothorax, pleural effusion, infiltrate, congestion, atelectasis, cardiomegaly, mass, and foreign object. Furthermore, we introduced an upper limit for the maximum waiting time, after which the highest urgency is assigned to the examination. RESULTS: The average RTAT for all critical findings was significantly reduced in all prioritization simulations compared to the FIFO simulation (e.g., pneumothorax: 35.6 min vs. 80.1 min; p < 0.0001), while the maximum RTAT for most findings increased at the same time (e.g., pneumothorax: 1293 min vs 890 min; p < 0.0001). Our "upper limit" substantially reduced the maximum RTAT in all classes (e.g., pneumothorax: 979 min vs. 1293 min/1178 min; p < 0.0001). CONCLUSION: Our simulations demonstrate that smart worklist prioritization by AI can reduce the average RTAT for critical findings in CXRs while maintaining a small maximum RTAT as FIFO. KEY POINTS: • Development of a realistic clinical workflow simulator based on empirical data from a hospital allowed precise assessment of smart worklist prioritization using artificial intelligence. • Employing a smart worklist prioritization without a threshold for maximum waiting time runs the risk of false negative predictions of the artificial intelligence greatly increasing the report turnaround time. • Use of a state-of-the-art convolution neural network can reduce the average report turnaround time almost to the upper limit of a perfect classification algorithm (e.g., pneumothorax: 35.6 min vs. 30.4 min).

Comparison of Deep Learning Approaches for Multi-Label Chest X-Ray Classification.

The increased availability of labeled X-ray image archives (e.g. ChestX-ray14 dataset) has triggered a growing interest in deep learning techniques. To provide better insight into the different approaches, and their applications to chest X-ray classification, we investigate a powerful network architecture in detail: the ResNet-50. Building on prior work in this domain, we consider transfer learning with and without fine-tuning as well as the training of a dedicated X-ray network from scratch. To leverage the high spatial resolution of X-ray data, we also include an extended ResNet-50 architecture, and a network integrating non-image data (patient age, gender and acquisition type) in the classification process. In a concluding experiment, we also investigate multiple ResNet depths (i.e. ResNet-38 and ResNet-101). In a systematic evaluation, using 5-fold re-sampling and a multi-label loss function, we compare the performance of the different approaches for pathology classification by ROC statistics and analyze differences between the classifiers using rank correlation. Overall, we observe a considerable spread in the achieved performance and conclude that the X-ray-specific ResNet-38, integrating non-image data yields the best overall results. Furthermore, class activation maps are used to understand the classification process, and a detailed analysis of the impact of non-image features is provided.

Optimal C-arm angulation during transcatheter aortic valve replacement: Accuracy of a rotational C-arm computed tomography based three dimensional heart model.

AIM: To investigate the accuracy of a rotational C-arm CT-based 3D heart model to predict an optimal C-arm configuration during transcatheter aortic valve replacement (TAVR). METHODS: Rotational C-arm CT (RCT) under rapid ventricular pacing was performed in 57 consecutive patients with severe aortic stenosis as part of the pre-procedural cardiac catheterization. With prototype software each RCT data set was segmented using a 3D heart model. From that the line of perpendicularity curve was obtained that generates a perpendicular view of the aortic annulus according to the right-cusp rule. To evaluate the accuracy of a model-based overlay we compared model- and expert-derived aortic root diameters. RESULTS: For all 57 patients in the RCT cohort diameter measurements were obtained from two independent operators and were compared to the model-based measurements. The inter-observer variability was measured to be in the range of 0°-12.96° of angular C-arm displacement for two independent operators. The model-to-operator agreement was 0°-13.82°. The model-based and expert measurements of aortic root diameters evaluated at the aortic annulus (r = 0.79, P < 0.01), the aortic sinus (r = 0.93, P < 0.01) and the sino-tubular junction (r = 0.92, P < 0.01) correlated on a high level and the Bland-Altman analysis showed good agreement. The interobserver measurements did not show a significant bias. CONCLUSION: Automatic segmentation of the aortic root using an anatomical model can accurately predict an optimal C-arm configuration, potentially simplifying current clinical workflows before and during TAVR.

A novel bone suppression method that improves lung nodule detection : Suppressing dedicated bone shadows in radiographs while preserving the remaining signal.

PURPOSE: Suppressing thoracic bone shadows in chest radiographs has been previously reported to improve the detection rates for solid lung nodules, however at the cost of increased false detection rates. These bone suppression methods are based on an artificial neural network that was trained using dual-energy subtraction images in order to mimic their appearance. METHOD: Here, a novel approach is followed where all bone shadows crossing the lung field are suppressed sequentially leaving the intercostal space unaffected. Given a contour delineating a bone, its image region is spatially transferred to separate normal image gradient components from tangential component. Smoothing the normal partial gradient along the contour results in a reconstruction of the image representing the bone shadow only, because all other overlaid signals tend to cancel out each other in this representation. RESULTS: The method works even with highly contrasted overlaid objects such as a pacemaker. The approach was validated in a reader study with two experienced chest radiologists, and these images helped improving both the sensitivity and the specificity of the readers for the detection and localization of solid lung nodules. The AUC improved significantly from 0.596 to 0.655 on a basis of 146 images from patients and normals with a total of 123 confirmed lung nodules. CONCLUSION: Subtracting all reconstructed bone shadows from the original image results in a soft image where lung nodules are no longer obscured by bone shadows. Both the sensitivity and the specificity of experienced radiologists increased.

Limited angle C-arm tomography and segmentation for guidance of atrial fibrillation ablation procedures.

Angiographic projections of the left atrium (LA) and the pulmonary veins (PV) acquired with a rotational C-arm system are used for 3D image reconstruction and subsequent automatic segmentation of the LA and PV to be used as roadmap in fluoroscopy guided LA ablation procedures. Acquisition of projections at high oblique angulations may be problematic due to increased collision danger of the detector with the right shoulder of the patient. We investigate the accuracy of image reconstruction and model based roadmap segmentation using limited angle C-arm tomography. The reduction of the angular range from 200 degrees to 150 degrees leads only to a moderate increase of the segmentation error from 1.5 mm to 2.0 mm if matched conditions are used in the segmentation, i.e., the model based segmentation is trained on images reconstructed with the same angular range as the test images. The minor decrease in accuracy may be outweighed by clinical workflow improvement, gained when large C-arm angulations can be avoided.

Automatic multi-model-based segmentation of the left atrium in cardiac MRI scans.

Model-based segmentation approaches have been proven to produce very accurate segmentation results while simultaneously providing an anatomic labeling for the segmented structures. However, variations of the anatomy, as they are often encountered e.g. on the drainage pattern of the pulmonary veins to the left atrium, cannot be represented by a single model. Automatic model selection extends the model-based segmentation approach to handling significant variational anatomies without user interaction. Using models for the three most common anatomical variations of the left atrium, we propose a method that uses an estimation of the local fit of different models to select the best fitting model automatically. Our approach employs the support vector machine for the automatic model selection. The method was evaluated on 42 very accurate segmentations of MRI scans using three different models. The correct model was chosen in 88.1% of the cases. In a second experiment, reflecting average segmentation results, the model corresponding to the clinical classification was automatically found in 78.0% of the cases.

Optimizing GHT-based heart localization in an automatic segmentation chain.

With automated image analysis tools entering rapidly the clinical practice, the demands regarding reliability, accuracy, and speed are strongly increasing. Systematic testing approaches to determine optimal parameter settings and to select algorithm design variants become essential in this context. We present an approach to optimize organ localization in a complex segmentation chain consisting of organ localization, parametric organ model adaptation, and deformable adaptation. In particular, we consider the Generalized Hough Transformation (GHT) and 3D heart segmentation in Computed Tomography Angiography (CTA) images. We rate the performance of our GHT variant by the initialization error and by computation time. Systematic parameter testing on a compute cluster allows to identify a parametrization with a good tradeoff between reliability and speed. This is achieved with coarse image sampling, a coarse Hough space resolution and a filtering step that we introduced to remove unspecific edges. Finally we show that optimization of the GHT parametrization results in a segmentation chain with reduced failure rates.

On the application of (topographic) independent and tree-dependent component analysis for the examination of DCE-MRI data.

In this contribution we investigate the applicability of different methods from the field of independent component analysis (ICA) for the examination of dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) data from breast cancer research. DCE-MRI has evolved in recent years as a powerful complement to X-ray based mammography for breast cancer diagnosis and monitoring. In DCE-MRI the time related development of the signal intensity after the administration of a contrast agent can provide valuable information about tissue states and characteristics. To this end, techniques related to ICA, offer promising options for data integration and feature extraction at voxel level. In order to evaluate the applicability of ICA, topographic ICA and tree-dependent component analysis (TCA), these methods are applied to twelve clinical cases from breast cancer research with a histopathologically confirmed diagnosis. For ICA these experiments are complemented by a reliability analysis of the estimated components. The outcome of all algorithms is quantitatively evaluated by means of receiver operating characteristics (ROC) statistics whereas the results for specific data sets are discussed exemplarily in terms of reification, score-plots and score images.

Libraries of synthetic stationary-phase and stress promoters as a tool for fine-tuning of expression of recombinant proteins in Escherichia coli.

Due to their induction characteristics stationary-phase promoters have a great potential in biotechnological processes for the production of heterologous proteins on a large-scale. In order to broaden the utility of stationary-phase promoters in bacterial expression systems and to create novel promoters induced by metabolic conditions, a library of synthetic stationary-phase/stress promoters for Escherichia coli was constructed. For designing the promoters the known -10 consensus sequence as well as the extended -10 region and an A/T-rich region downstream of the -10 region were kept constant, while sequences from -37 to -14 were partially or completely randomized. For detection and selection of stationary-phase promoters GFP with enhanced fluorescence was used. The expression pattern of the GFP reporter system was compared with that of the LacZ reporter system. To screen and characterize colonies containing stationary-phase/stress promoters a bioinformatic approach was developed. In total, 33 promoters were selected which cover a broad range of promoter activities and induction times indicating that the strength of promoters can be modulated by partially randomizing the sequence upstream of the -10 region. The induction ratio of synthetic promoters at the transition from exponential to stationary-phase was from 4 to over 6000 and the induction time relative to the entrance into stationary-phase from -1.4 to 2.7 h. Ninety-one percentage of the promoters had no or only low background activity during exponential growth. The broad variability of the promoters offers good possibilities for fine-tuning of gene expression and for applications in industrial bioprocesses.

Reagent-free automatic cell viability determination using neural networks based machine vision and dark-field microscopy in Saccharomyces cerevisiae.

Fermentation industries require in-situ real-time monitoring of cell viability during fermentation processes. For this purpose, reagent-free approaches are desired because they can be used for in situ analysis and reduce the system's complexity. We have developed an automatic way of determining cell viability via analysis of time-lapse image sequences taken by dark field microscopy without the aid of any additional reagents. The image processing is based on neural networks based machine vision, involving Principal Component Analysis (PCA) to investigate the dynamic information of intracellular movements. In consequence, the essential features as the vital sign of the target cells are discovered. Viability predictions using the Support Vector Machine (SVM) classifier have been done successfully on the datasets with different qualities. Accuracy up to above 90% has been obtained on the basis of image enhancement. Robustness of the system is proved by the results of the tests. The model organism we have used is Saccharomyces cerevisiae, however, this technique can promisingly be applied for the identification of cell viability of other organisms as well.

Image fusion for dynamic contrast enhanced magnetic resonance imaging.

BACKGROUND: Multivariate imaging techniques such as dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) have been shown to provide valuable information for medical diagnosis. Even though these techniques provide new information, integrating and evaluating the much wider range of information is a challenging task for the human observer. This task may be assisted with the use of image fusion algorithms. METHODS: In this paper, image fusion based on Kernel Principal Component Analysis (KPCA) is proposed for the first time. It is demonstrated that a priori knowledge about the data domain can be easily incorporated into the parametrisation of the KPCA, leading to task-oriented visualisations of the multivariate data. The results of the fusion process are compared with those of the well-known and established standard linear Principal Component Analysis (PCA) by means of temporal sequences of 3D MRI volumes from six patients who took part in a breast cancer screening study. RESULTS: The PCA and KPCA algorithms are able to integrate information from a sequence of MRI volumes into informative gray value or colour images. By incorporating a priori knowledge, the fusion process can be automated and optimised in order to visualise suspicious lesions with high contrast to normal tissue. CONCLUSION: Our machine learning based image fusion approach maps the full signal space of a temporal DCE-MRI sequence to a single meaningful visualisation with good tissue/lesion contrast and thus supports the radiologist during manual image evaluation.