Self-supervised learning for interventional image analytics: toward robust device trackers.

Purpose: The accurate detection and tracking of devices, such as guiding catheters in live X-ray image acquisitions, are essential prerequisites for endovascular cardiac interventions. This information is leveraged for procedural guidance, e.g., directing stent placements. To ensure procedural safety and efficacy, there is a need for high robustness/no failures during tracking. To achieve this, one needs to efficiently tackle challenges, such as device obscuration by the contrast agent or other external devices or wires and changes in the field-of-view or acquisition angle, as well as the continuous movement due to cardiac and respiratory motion. Approach: To overcome the aforementioned challenges, we propose an approach to learn spatio-temporal features from a very large data cohort of over 16 million interventional X-ray frames using self-supervision for image sequence data. Our approach is based on a masked image modeling technique that leverages frame interpolation-based reconstruction to learn fine inter-frame temporal correspondences. The features encoded in the resulting model are fine-tuned downstream in a light-weight model. Results: Our approach achieves state-of-the-art performance, in particular for robustness, compared to ultra optimized reference solutions (that use multi-stage feature fusion or multi-task and flow regularization). The experiments show that our method achieves a 66.31% reduction in the maximum tracking error against the reference solutions (23.20% when flow regularization is used), achieving a success score of 97.95% at a 3× faster inference speed of 42 frames-per-second (on GPU). In addition, we achieve a 20% reduction in the standard deviation of errors, which indicates a much more stable tracking performance. Conclusions: The proposed data-driven approach achieves superior performance, particularly in robustness and speed compared with the frequently used multi-modular approaches for device tracking. The results encourage the use of our approach in various other tasks within interventional image analytics that require effective understanding of spatio-temporal semantics.

Large-Scale Study on AI's Impact on Identifying Chest Radiographs with No Actionable Disease in Outpatient Imaging.

RATIONALE AND OBJECTIVES: Given the high volume of chest radiographs, radiologists frequently encounter heavy workloads. In outpatient imaging, a substantial portion of chest radiographs show no actionable findings. Automatically identifying these cases could improve efficiency by facilitating shorter reading workflows. PURPOSE: A large-scale study to assess the performance of AI on identifying chest radiographs with no actionable disease (NAD) in an outpatient imaging population using comprehensive, objective, and reproducible criteria for NAD. MATERIALS AND METHODS: The independent validation study includes 15000 patients with chest radiographs in posterior-anterior (PA) and lateral projections from an outpatient imaging center in the United States. Ground truth was established by reviewing CXR reports and classifying cases as NAD or actionable disease (AD). The NAD definition includes completely normal chest radiographs and radiographs with well-defined non-actionable findings. The AI NAD Analyzer1 (trained with 100 million multimodal images and fine-tuned on 1.3 million radiographs) utilizes a tandem system with image-level rule in and compartment-level rule out to provide case level output as NAD or potential actionable disease (PAD). RESULTS: A total of 14057 cases met our eligibility criteria (age 56 ± 16.1 years, 55% women and 45% men). The prevalence of NAD cases in the study population was 70.7%. The AI NAD Analyzer correctly classified NAD cases with a sensitivity of 29.1% and a yield of 20.6%. The specificity was 98.9% which corresponds to a miss rate of 0.3% of cases. Significant findings were missed in 0.06% of cases, while no cases with critical findings were missed by AI. CONCLUSION: In an outpatient population, AI can identify 20% of chest radiographs as NAD with a very low rate of missed findings. These cases could potentially be read using a streamlined protocol, thus improving efficiency and consequently reducing daily workload for radiologists.

Cardiac ultrasound simulation for autonomous ultrasound navigation.

Introduction: Ultrasound is well-established as an imaging modality for diagnostic and interventional purposes. However, the image quality varies with operator skills as acquiring and interpreting ultrasound images requires extensive training due to the imaging artefacts, the range of acquisition parameters and the variability of patient anatomies. Automating the image acquisition task could improve acquisition reproducibility and quality but training such an algorithm requires large amounts of navigation data, not saved in routine examinations. Methods: We propose a method to generate large amounts of ultrasound images from other modalities and from arbitrary positions, such that this pipeline can later be used by learning algorithms for navigation. We present a novel simulation pipeline which uses segmentations from other modalities, an optimized volumetric data representation and GPU-accelerated Monte Carlo path tracing to generate view-dependent and patient-specific ultrasound images. Results: We extensively validate the correctness of our pipeline with a phantom experiment, where structures' sizes, contrast and speckle noise properties are assessed. Furthermore, we demonstrate its usability to train neural networks for navigation in an echocardiography view classification experiment by generating synthetic images from more than 1,000 patients. Networks pre-trained with our simulations achieve significantly superior performance in settings where large real datasets are not available, especially for under-represented classes. Discussion: The proposed approach allows for fast and accurate patient-specific ultrasound image generation, and its usability for training networks for navigation-related tasks is demonstrated.

Separable Tendon-Driven Robotic Manipulator with a Long, Flexible, Passive Proximal Section.

This work tackles practical issues which arise when using a tendon-driven robotic manipulator (TDRM) with a long, flexible, passive proximal section in medical applications. Tendon-driven devices are preferred in medicine for their improved outcomes via minimally invasive procedures, but TDRMs come with unique challenges such as sterilization and reuse, simultaneous control of tendons, hysteresis in the tendon-sheath mechanism, and unmodeled effects of the proximal section shape. A separable TDRM which overcomes difficulties in actuation and sterilization is introduced, in which the body containing the electronics is reusable and the remainder is disposable. An open-loop redundant controller which resolves the redundancy in the kinematics is developed. Simple linear hysteresis compensation and re-tension compensation based on the physical properties of the device are proposed. The controller and compensation methods are evaluated on a testbed for a straight proximal section, a curved proximal section at various static angles, and a proximal section which dynamically changes angles; and overall, distal tip error was reduced.

AUCReshaping: improved sensitivity at high-specificity.

The evaluation of deep-learning (DL) systems typically relies on the Area under the Receiver-Operating-Curve (AU-ROC) as a performance metric. However, AU-ROC, in its holistic form, does not sufficiently consider performance within specific ranges of sensitivity and specificity, which are critical for the intended operational context of the system. Consequently, two systems with identical AU-ROC values can exhibit significantly divergent real-world performance. This issue is particularly pronounced in the context of anomaly detection tasks, a commonly employed application of DL systems across various research domains, including medical imaging, industrial automation, manufacturing, cyber security, fraud detection, and drug research, among others. The challenge arises from the heavy class imbalance in training datasets, with the abnormality class often incurring a considerably higher misclassification cost compared to the normal class. Traditional DL systems address this by adjusting the weighting of the cost function or optimizing for specific points along the ROC curve. While these approaches yield reasonable results in many cases, they do not actively seek to maximize performance for the desired operating point. In this study, we introduce a novel technique known as AUCReshaping, designed to reshape the ROC curve exclusively within the specified sensitivity and specificity range, by optimizing sensitivity at a predetermined specificity level. This reshaping is achieved through an adaptive and iterative boosting mechanism that allows the network to focus on pertinent samples during the learning process. We primarily investigated the impact of AUCReshaping in the context of abnormality detection tasks, specifically in Chest X-Ray (CXR) analysis, followed by breast mammogram and credit card fraud detection tasks. The results reveal a substantial improvement, ranging from 2 to 40%, in sensitivity at high-specificity levels for binary classification tasks.

Contrastive self-supervised learning from 100 million medical images with optional supervision.

Purpose: Building accurate and robust artificial intelligence systems for medical image assessment requires the creation of large sets of annotated training examples. However, constructing such datasets is very costly due to the complex nature of annotation tasks, which often require expert knowledge (e.g., a radiologist). To counter this limitation, we propose a method to learn from medical images at scale in a self-supervised way. Approach: Our approach, based on contrastive learning and online feature clustering, leverages training datasets of over 100,000,000 medical images of various modalities, including radiography, computed tomography (CT), magnetic resonance (MR) imaging, and ultrasonography (US). We propose to use the learned features to guide model training in supervised and hybrid self-supervised/supervised regime on various downstream tasks. Results: We highlight a number of advantages of this strategy on challenging image assessment problems in radiography, CT, and MR: (1) significant increase in accuracy compared to the state-of-the-art (e.g., area under the curve boost of 3% to 7% for detection of abnormalities from chest radiography scans and hemorrhage detection on brain CT); (2) acceleration of model convergence during training by up to 85% compared with using no pretraining (e.g., 83% when training a model for detection of brain metastases in MR scans); and (3) increase in robustness to various image augmentations, such as intensity variations, rotations or scaling reflective of data variation seen in the field. Conclusions: The proposed approach enables large gains in accuracy and robustness on challenging image assessment problems. The improvement is significant compared with other state-of-the-art approaches trained on medical or vision images (e.g., ImageNet).

Value of quantitative airspace disease measured on chest CT and chest radiography at initial diagnosis compared to clinical variables for prediction of severe COVID-19.

Purpose: Rapid prognostication of COVID-19 patients is important for efficient resource allocation. We evaluated the relative prognostic value of baseline clinical variables (CVs), quantitative human-read chest CT (qCT), and AI-read chest radiograph (qCXR) airspace disease (AD) in predicting severe COVID-19. Approach: We retrospectively selected 131 COVID-19 patients (SARS-CoV-2 positive, March to October, 2020) at a tertiary hospital in the United States, who underwent chest CT and CXR within 48 hr of initial presentation. CVs included patient demographics and laboratory values; imaging variables included qCT volumetric percentage AD (POv) and qCXR area-based percentage AD (POa), assessed by a deep convolutional neural network. Our prognostic outcome was need for ICU admission. We compared the performance of three logistic regression models: using CVs known to be associated with prognosis (model I), using a dimension-reduced set of best predictor variables (model II), and using only age and AD (model III). Results: 60/131 patients required ICU admission, whereas 71/131 did not. Model I performed the poorest ( AUC = 0.67 [0.58 to 0.76]; accuracy = 77 % ). Model II performed the best ( AUC = 0.78 [0.71 to 0.86]; accuracy = 81 % ). Model III was equivalent ( AUC = 0.75 [0.67 to 0.84]; accuracy = 80 % ). Both models II and III outperformed model I ( AUC difference = 0.11 [0.02 to 0.19], p = 0.01 ; AUC difference = 0.08 [0.01 to 0.15], p = 0.04 , respectively). Model II and III results did not change significantly when POv was replaced by POa. Conclusions: Severe COVID-19 can be predicted using only age and quantitative AD imaging metrics at initial diagnosis, which outperform the set of CVs. Moreover, AI-read qCXR can replace qCT metrics without loss of prognostic performance, promising more resource-efficient prognostication.

Robust classification from noisy labels: Integrating additional knowledge for chest radiography abnormality assessment.

Chest radiography is the most common radiographic examination performed in daily clinical practice for the detection of various heart and lung abnormalities. The large amount of data to be read and reported, with more than 100 studies per day for a single radiologist, poses a challenge in consistently maintaining high interpretation accuracy. The introduction of large-scale public datasets has led to a series of novel systems for automated abnormality classification. However, the labels of these datasets were obtained using natural language processed medical reports, yielding a large degree of label noise that can impact the performance. In this study, we propose novel training strategies that handle label noise from such suboptimal data. Prior label probabilities were measured on a subset of training data re-read by 4 board-certified radiologists and were used during training to increase the robustness of the training model to the label noise. Furthermore, we exploit the high comorbidity of abnormalities observed in chest radiography and incorporate this information to further reduce the impact of label noise. Additionally, anatomical knowledge is incorporated by training the system to predict lung and heart segmentation, as well as spatial knowledge labels. To deal with multiple datasets and images derived from various scanners that apply different post-processing techniques, we introduce a novel image normalization strategy. Experiments were performed on an extensive collection of 297,541 chest radiographs from 86,876 patients, leading to a state-of-the-art performance level for 17 abnormalities from 2 datasets. With an average AUC score of 0.880 across all abnormalities, our proposed training strategies can be used to significantly improve performance scores.

Automated Detection and Quantification of COVID-19 Airspace Disease on Chest Radiographs: A Novel Approach Achieving Expert Radiologist-Level Performance Using a Deep Convolutional Neural Network Trained on Digital Reconstructed Radiographs From Computed Tomography-Derived Ground Truth.

OBJECTIVES: The aim of this study was to leverage volumetric quantification of airspace disease (AD) derived from a superior modality (computed tomography [CT]) serving as ground truth, projected onto digitally reconstructed radiographs (DRRs) to (1) train a convolutional neural network (CNN) to quantify AD on paired chest radiographs (CXRs) and CTs, and (2) compare the DRR-trained CNN to expert human readers in the CXR evaluation of patients with confirmed COVID-19. MATERIALS AND METHODS: We retrospectively selected a cohort of 86 COVID-19 patients (with positive reverse transcriptase-polymerase chain reaction test results) from March to May 2020 at a tertiary hospital in the northeastern United States, who underwent chest CT and CXR within 48 hours. The ground-truth volumetric percentage of COVID-19-related AD (POv) was established by manual AD segmentation on CT. The resulting 3-dimensional masks were projected into 2-dimensional anterior-posterior DRR to compute area-based AD percentage (POa). A CNN was trained with DRR images generated from a larger-scale CT dataset of COVID-19 and non-COVID-19 patients, automatically segmenting lungs, AD, and quantifying POa on CXR. The CNN POa results were compared with POa quantified on CXR by 2 expert readers and to the POv ground truth, by computing correlations and mean absolute errors. RESULTS: Bootstrap mean absolute error and correlations between POa and POv were 11.98% (11.05%-12.47%) and 0.77 (0.70-0.82) for average of expert readers and 9.56% to 9.78% (8.83%-10.22%) and 0.78 to 0.81 (0.73-0.85) for the CNN, respectively. CONCLUSIONS: Our CNN trained with DRR using CT-derived airspace quantification achieved expert radiologist level of accuracy in the quantification of AD on CXR in patients with positive reverse transcriptase-polymerase chain reaction test results for COVID-19.

No Surprises: Training Robust Lung Nodule Detection for Low-Dose CT Scans by Augmenting With Adversarial Attacks.

Detecting malignant pulmonary nodules at an early stage can allow medical interventions which may increase the survival rate of lung cancer patients. Using computer vision techniques to detect nodules can improve the sensitivity and the speed of interpreting chest CT for lung cancer screening. Many studies have used CNNs to detect nodule candidates. Though such approaches have been shown to outperform the conventional image processing based methods regarding the detection accuracy, CNNs are also known to be limited to generalize on under-represented samples in the training set and prone to imperceptible noise perturbations. Such limitations can not be easily addressed by scaling up the dataset or the models. In this work, we propose to add adversarial synthetic nodules and adversarial attack samples to the training data to improve the generalization and the robustness of the lung nodule detection systems. To generate hard examples of nodules from a differentiable nodule synthesizer, we use projected gradient descent (PGD) to search the latent code within a bounded neighbourhood that would generate nodules to decrease the detector response. To make the network more robust to unanticipated noise perturbations, we use PGD to search for noise patterns that can trigger the network to give over-confident mistakes. By evaluating on two different benchmark datasets containing consensus annotations from three radiologists, we show that the proposed techniques can improve the detection performance on real CT data. To understand the limitations of both the conventional networks and the proposed augmented networks, we also perform stress-tests on the false positive reduction networks by feeding different types of artificially produced patches. We show that the augmented networks are more robust to both under-represented nodules as well as resistant to noise perturbations.

Quantifying and leveraging predictive uncertainty for medical image assessment.

The interpretation of medical images is a challenging task, often complicated by the presence of artifacts, occlusions, limited contrast and more. Most notable is the case of chest radiography, where there is a high inter-rater variability in the detection and classification of abnormalities. This is largely due to inconclusive evidence in the data or subjective definitions of disease appearance. An additional example is the classification of anatomical views based on 2D Ultrasound images. Often, the anatomical context captured in a frame is not sufficient to recognize the underlying anatomy. Current machine learning solutions for these problems are typically limited to providing probabilistic predictions, relying on the capacity of underlying models to adapt to limited information and the high degree of label noise. In practice, however, this leads to overconfident systems with poor generalization on unseen data. To account for this, we propose a system that learns not only the probabilistic estimate for classification, but also an explicit uncertainty measure which captures the confidence of the system in the predicted output. We argue that this approach is essential to account for the inherent ambiguity characteristic of medical images from different radiologic exams including computed radiography, ultrasonography and magnetic resonance imaging. In our experiments we demonstrate that sample rejection based on the predicted uncertainty can significantly improve the ROC-AUC for various tasks, e.g., by 8% to 0.91 with an expected rejection rate of under 25% for the classification of different abnormalities in chest radiographs. In addition, we show that using uncertainty-driven bootstrapping to filter the training data, one can achieve a significant increase in robustness and accuracy. Finally, we present a multi-reader study showing that the predictive uncertainty is indicative of reader errors.

Towards intelligent robust detection of anatomical structures in incomplete volumetric data.

Robust and fast detection of anatomical structures represents an important component of medical image analysis technologies. Current solutions for anatomy detection are based on machine learning, and are generally driven by suboptimal and exhaustive search strategies. In particular, these techniques do not effectively address cases of incomplete data, i.e., scans acquired with a partial field-of-view. We address these challenges by following a new paradigm, which reformulates the detection task to teaching an intelligent artificial agent how to actively search for an anatomical structure. Using the principles of deep reinforcement learning with multi-scale image analysis, artificial agents are taught optimal navigation paths in the scale-space representation of an image, while accounting for structures that are missing from the field-of-view. The spatial coherence of the observed anatomical landmarks is ensured using elements from statistical shape modeling and robust estimation theory. Experiments show that our solution outperforms marginal space deep learning, a powerful deep learning method, at detecting different anatomical structures without any failure. The dataset contains 5043 3D-CT volumes from over 2000 patients, totaling over 2,500,000 image slices. In particular, our solution achieves 0% false-positive and 0% false-negative rates at detecting whether the landmarks are captured in the field-of-view of the scan (excluding all border cases), with an average detection accuracy of 2.78 mm. In terms of runtime, we reduce the detection-time of the marginal space deep learning method by 20-30 times to under 40 ms, an unmatched performance for high resolution incomplete 3D-CT data.

Marginal Space Deep Learning: Efficient Architecture for Volumetric Image Parsing.

Robust and fast solutions for anatomical object detection and segmentation support the entire clinical workflow from diagnosis, patient stratification, therapy planning, intervention and follow-up. Current state-of-the-art techniques for parsing volumetric medical image data are typically based on machine learning methods that exploit large annotated image databases. Two main challenges need to be addressed, these are the efficiency in scanning high-dimensional parametric spaces and the need for representative image features which require significant efforts of manual engineering. We propose a pipeline for object detection and segmentation in the context of volumetric image parsing, solving a two-step learning problem: anatomical pose estimation and boundary delineation. For this task we introduce Marginal Space Deep Learning (MSDL), a novel framework exploiting both the strengths of efficient object parametrization in hierarchical marginal spaces and the automated feature design of Deep Learning (DL) network architectures. In the 3D context, the application of deep learning systems is limited by the very high complexity of the parametrization. More specifically 9 parameters are necessary to describe a restricted affine transformation in 3D, resulting in a prohibitive amount of billions of scanning hypotheses. The mechanism of marginal space learning provides excellent run-time performance by learning classifiers in clustered, high-probability regions in spaces of gradually increasing dimensionality. To further increase computational efficiency and robustness, in our system we learn sparse adaptive data sampling patterns that automatically capture the structure of the input. Given the object localization, we propose a DL-based active shape model to estimate the non-rigid object boundary. Experimental results are presented on the aortic valve in ultrasound using an extensive dataset of 2891 volumes from 869 patients, showing significant improvements of up to 45.2% over the state-of-the-art. To our knowledge, this is the first successful demonstration of the DL potential to detection and segmentation in full 3D data with parametrized representations.