Search

Detection Transformers represent end-to-end object detection approaches based on a Transformer encoder-decoder architecture, exploiting the attention mechanism for global relation modeling. Although Detection Transformers deliver results on par with or even superior to their highly optimized CNN-based counterparts operating on 2D natural images, their success is closely coupled to access to a vast amount of training data. This, however, restricts the feasibility of employing Detection Transformers in the medical domain, as access to annotated data is typically limited. To tackle this issue and facilitate the advent of medical Detection Transformers, we propose a novel Detection Transformer for 3D anatomical structure detection, dubbed Focused Decoder. Focused Decoder leverages information from an anatomical region atlas to simultaneously deploy query anchors and restrict the cross-attention's field of view to regions of interest, which allows for a precise focus on relevant anatomical structures. We evaluate our proposed approach on two publicly available CT datasets and demonstrate that Focused Decoder not only provides strong detection results and thus alleviates the need for a vast amount of annotated data but also exhibits exceptional and highly intuitive explainability of results via attention weights. Our code is available at https://github.com/bwittmann/transoar., Accepted for publication at the Journal of Machine Learning for Biomedical Imaging (MELBA) https://melba-journal.org/2023:003

Collateral circulation results from specialized anastomotic channels which are capable of providing oxygenated blood to regions with compromised blood flow caused by arterial obstruction. The quality of collateral circulation has been established as a key factor in determining the likelihood of a favorable clinical outcome and goes a long way to determining the choice of a stroke care model. Though many imaging and grading methods exist for quantifying collateral blood flow, the actual grading is mostly done through manual inspection. This approach is associated with a number of challenges. First, it is time-consuming. Second, there is a high tendency for bias and inconsistency in the final grade assigned to a patient depending on the experience level of the clinician. We present a multi-stage deep learning approach to predict collateral flow grading in stroke patients based on radiomic features extracted from MR perfusion data. First, we formulate a region of interest detection task as a reinforcement learning problem and train a deep learning network to automatically detect the occluded region within the 3D MR perfusion volumes. Second, we extract radiomic features from the obtained region of interest through local image descriptors and denoising auto-encoders. Finally, we apply a convolutional neural network and other machine learning classifiers to the extracted radiomic features to automatically predict the collateral flow grading of the given patient volume as one of three severity classes - no flow (0), moderate flow (1), and good flow (2). Results from our experiments show an overall accuracy of 72% in the three-class prediction task. With an inter-observer agreement of 16% and a maximum intra-observer agreement of 74% in a similar experiment, our automated deep learning approach demonstrates a performance comparable to expert grading, is faster than visual inspection, and eliminates the problem of grading bias.

Gliomas are the most common type of primary brain tumors. Although gliomas are relatively rare, they are among the deadliest types of cancer, with a survival rate of less than 2 years after diagnosis. Gliomas are challenging to diagnose, hard to treat and inherently resistant to conventional therapy. Years of extensive research to improve diagnosis and treatment of gliomas have decreased mortality rates across the Global North, while chances of survival among individuals in low- and middle-income countries (LMICs) remain unchanged and are significantly worse in Sub-Saharan Africa (SSA) populations. Long-term survival with glioma is associated with the identification of appropriate pathological features on brain MRI and confirmation by histopathology. Since 2012, the Brain Tumor Segmentation (BraTS) Challenge have evaluated state-of-the-art machine learning methods to detect, characterize, and classify gliomas. However, it is unclear if the state-of-the-art methods can be widely implemented in SSA given the extensive use of lower-quality MRI technology, which produces poor image contrast and resolution and more importantly, the propensity for late presentation of disease at advanced stages as well as the unique characteristics of gliomas in SSA (i.e., suspected higher rates of gliomatosis cerebri). Thus, the BraTS-Africa Challenge provides a unique opportunity to include brain MRI glioma cases from SSA in global efforts through the BraTS Challenge to develop and evaluate computer-aided-diagnostic (CAD) methods for the detection and characterization of glioma in resource-limited settings, where the potential for CAD tools to transform healthcare are more likely., Comment: arXiv admin note: text overlap with arXiv:2107.02314

Clinical monitoring of metastatic disease to the brain can be a laborious and time-consuming process, especially in cases involving multiple metastases when the assessment is performed manually. The Response Assessment in Neuro-Oncology Brain Metastases (RANO-BM) guideline, which utilizes the unidimensional longest diameter, is commonly used in clinical and research settings to evaluate response to therapy in patients with brain metastases. However, accurate volumetric assessment of the lesion and surrounding peri-lesional edema holds significant importance in clinical decision-making and can greatly enhance outcome prediction. The unique challenge in performing segmentations of brain metastases lies in their common occurrence as small lesions. Detection and segmentation of lesions that are smaller than 10 mm in size has not demonstrated high accuracy in prior publications. The brain metastases challenge sets itself apart from previously conducted MICCAI challenges on glioma segmentation due to the significant variability in lesion size. Unlike gliomas, which tend to be larger on presentation scans, brain metastases exhibit a wide range of sizes and tend to include small lesions. We hope that the BraTS-METS dataset and challenge will advance the field of automated brain metastasis detection and segmentation.

Registration methods facilitate the comparison of multiparametric magnetic resonance images acquired at different stages of brain tumor treatments. Image-based registration solutions are influenced by the sequences chosen to compute the distance measure, and the lack of image correspondences due to the resection cavities and pathological tissues. Nonetheless, an evaluation of the impact of these input parameters on the registration of longitudinal data is still missing. This work evaluates the influence of multiple sequences, namely T1-weighted (T1), T2-weighted (T2), contrast enhanced T1-weighted (T1-CE), and T2 Fluid Attenuated Inversion Recovery (FLAIR), and the exclusion of the pathological tissues on the non-rigid registration of pre- and post-operative images. We here investigate two types of registration methods, an iterative approach and a convolutional neural network solution based on a 3D U-Net. We employ two test sets to compute the mean target registration error (mTRE) based on corresponding landmarks. In the first set, markers are positioned exclusively in the surroundings of the pathology. The methods employing T1-CE achieves the lowest mTREs, with a improvement up to 0.8 mm for the iterative solution. The results are higher than the baseline when using the FLAIR sequence. When excluding the pathology, lower mTREs are observable for most of the methods. In the second test set, corresponding landmarks are located in the entire brain volumes. Both solutions employing T1-CE obtain the lowest mTREs, with a decrease up to 1.16 mm for the iterative method, whereas the results worsen using the FLAIR. When excluding the pathology, an improvement is observable for the CNN method using T1-CE. Both approaches utilizing the T1-CE sequence obtain the best mTREs, whereas the FLAIR is the least informative to guide the registration process. Besides, the exclusion of pathology from the distance measure computation improves the registration of the brain tissues surrounding the tumor. Thus, this work provides the first numerical evaluation of the influence of these parameters on the registration of longitudinal magnetic resonance images, and it can be helpful for developing future algorithms.

Exploiting 4D-flow magnetic resonance imaging (MRI) data to quantify hemodynamics requires an adequate spatio-temporal vector field resolution at a low noise level. To address this challenge, we provide a learned solution to super-resolve in vivo 4D-flow MRI data at a post-processing level. We propose a deep convolutional neural network (CNN) that learns the inter-scale relationship of the velocity vector map and leverages an efficient residual learning scheme to make it computationally feasible. A novel, direction-sensitive, and robust loss function is crucial to learning vector-field data. We present a detailed comparative study between the proposed super-resolution and the conventional cubic B-spline based vector-field super-resolution. Our method improves the peak-velocity to noise ratio of the flow field by 10 and 30% for in vivo cardiovascular and cerebrovascular data, respectively, for 4 × super-resolution over the state-of-the-art cubic B-spline. Significantly, our method offers 10x faster inference over the cubic B-spline. The proposed approach for super-resolution of 4D-flow data would potentially improve the subsequent calculation of hemodynamic quantities.

Automatic localization and segmentation of organs-at-risk (OAR) in CT are essential pre-processing steps in medical image analysis tasks, such as radiation therapy planning. For instance, the segmentation of OAR surrounding tumors enables the maximization of radiation to the tumor area without compromising the healthy tissues. However, the current medical workflow requires manual delineation of OAR, which is prone to errors and is annotator-dependent. In this work, we aim to introduce a unified 3D pipeline for OAR localization-segmentation rather than novel localization or segmentation architectures. To the best of our knowledge, our proposed framework fully enables the exploitation of 3D context information inherent in medical imaging. In the first step, a 3D multi-variate regression network predicts organs' centroids and bounding boxes. Secondly, 3D organ-specific segmentation networks are leveraged to generate a multi-organ segmentation map. Our method achieved an overall Dice score of $0.9260\pm 0.18 \%$ on the VISCERAL dataset containing CT scans with varying fields of view and multiple organs.

Self-supervised pretext tasks have been introduced as an effective strategy when learning target tasks on small annotated data sets. However, while current research focuses on exploring novel pretext tasks for meaningful and reusable representation learning for the target task, the study of its robustness and generalizability has remained relatively under-explored. Specifically, it is crucial in medical imaging to proactively investigate performance under different perturbations for reliable deployment of clinical applications. In this work, we revisit medical imaging networks pre-trained with self-supervised learnings and categorically evaluate robustness and generalizability compared to vanilla supervised learning. Our experiments on pneumonia detection in X-rays and multi-organ segmentation in CT yield conclusive results exposing the hidden benefits of self-supervision pre-training for learning robust feature representations.

Designing of a cranial implant needs a 3D understanding of the complete skull shape. Thus, taking a 2D approach is sub-optimal, since a 2D model lacks a holistic 3D view of both the defective and healthy skulls. Further, loading the whole 3D skull shapes at its original image resolution is not feasible in commonly available GPUs. To mitigate these issues, we propose a fully convolutional network composed of two subnetworks. The first subnetwork is designed to complete the shape of the downsampled defective skull. The second subnetwork upsamples the reconstructed shape slice-wise. We train the 3D and 2D networks together end-to-end, with a hierarchical loss function. Our proposed solution accurately predicts a high-resolution 3D implant in the challenge test case in terms of dice-score and the Hausdorff distance.

Osteoporotic vertebral fractures have a severe impact on patients' overall well-being but are severely under-diagnosed. These fractures present themselves at various levels of severity measured using the Genant's grading scale. Insufficient annotated datasets, severe data-imbalance, and minor difference in appearances between fractured and healthy vertebrae make naive classification approaches result in poor discriminatory performance. Addressing this, we propose a representation learning-inspired approach for automated vertebral fracture detection, aimed at learning latent representations efficient for fracture detection. Building on state-of-art metric losses, we present a novel Grading Loss for learning representations that respect Genant's fracture grading scheme. On a publicly available spine dataset, the proposed loss function achieves a fracture detection F1 score of 81.5%, a 10% increase over a naive classification baseline., Comment: To be presented at MICCAI 2020

Mammogram is the most effective imaging modality for the mass lesion detection of breast cancer at the early stage. The information from the two paired views (i.e., medio-lateral oblique and cranio-caudal) are highly relational and complementary, and this is crucial for doctors' decisions in clinical practice. However, existing mass detection methods do not consider jointly learning effective features from the two relational views. To address this issue, this paper proposes a novel mammogram mass detection framework, termed Cross-View Relation Region-based Convolutional Neural Networks (CVR-RCNN). The proposed CVR-RCNN is expected to capture the latent relation information between the corresponding mass region of interests (ROIs) from the two paired views. Evaluations on a new large-scale private dataset and a public mammogram dataset show that the proposed CVR-RCNN outperforms existing state-of-the-art mass detection methods. Meanwhile, our experimental results suggest that incorporating the relation information across two views helps to train a superior detection model, which is a promising avenue for mammogram mass detection.

Early diagnosis of pulmonary nodules (PNs) can improve the survival rate of patients and yet is a challenging task for radiologists due to the image noise and artifacts in computed tomography (CT) images. In this paper, we propose a novel and effective abnormality detector implementing the attention mechanism and group convolution on 3D single-shot detector (SSD) called group-attention SSD (GA-SSD). We find that group convolution is effective in extracting rich context information between continuous slices, and attention network can learn the target features automatically. We collected a large-scale dataset that contained 4146 CT scans with annotations of varying types and sizes of PNs (even PNs smaller than 3mm were annotated). To the best of our knowledge, this dataset is the largest cohort with relatively complete annotations for PNs detection. Our experimental results show that the proposed group-attention SSD outperforms the classic SSD framework as well as the state-of-the-art 3DCNN, especially on some challenging lesion types.

Arterial spin labeling (ASL) allows to quantify the cerebral blood flow (CBF) by magnetic labeling of the arterial blood water. ASL is increasingly used in clinical studies due to its noninvasiveness, repeatability and benefits in quantification. However, ASL suffers from an inherently low-signal-to-noise ratio (SNR) requiring repeated measurements of control/spin-labeled (C/L) pairs to achieve a reasonable image quality, which in return increases motion sensitivity. This leads to clinically prolonged scanning times increasing the risk of motion artifacts. Thus, there is an immense need of advanced imaging and processing techniques in ASL. In this paper, we propose a novel deep learning based approach to improve the perfusion-weighted image quality obtained from a subset of all available pairwise C/L subtractions. Specifically, we train a deep fully convolutional network (FCN) to learn a mapping from noisy perfusion-weighted image and its subtraction (residual) from the clean image. Additionally, we incorporate the CBF estimation model in the loss function during training, which enables the network to produce high quality images while simultaneously enforcing the CBF estimates to be as close as reference CBF values. Extensive experiments on synthetic and clinical ASL datasets demonstrate the effectiveness of our method in terms of improved ASL image quality, accurate CBF parameter estimation and considerably small computation time during testing., Comment: Accepted at MICCAI 2018. 9 pages, 4 figures, 1 table

The segmentation of liver lesions is crucial for detection, diagnosis and monitoring progression of liver cancer. However, design of accurate automated methods remains challenging due to high noise in CT scans, low contrast between liver and lesions, as well as large lesion variability. We propose a 3D automatic, unsupervised method for liver lesions segmentation using a phase separation approach. It is assumed that liver is a mixture of two phases: healthy liver and lesions, represented by different image intensities polluted by noise. The Cahn-Hilliard equation is used to remove the noise and separate the mixture into two distinct phases with well-defined interfaces. This simplifies the lesion detection and segmentation task drastically and enables to segment liver lesions by thresholding the Cahn-Hilliard solution. The method was tested on 3Dircadb and LITS dataset.

The accurate localization of anatomical landmarks is a challenging task, often solved by domain specific approaches. We propose a method for the automatic localization of landmarks in complex, repetitive anatomical structures. The key idea is to combine three steps: (1) a classifier for pre-filtering anatomical landmark positions that (2) are refined through a Hough regression model, together with (3) a parts-based model of the global landmark topology to select the final landmark positions. During training landmarks are annotated in a set of example volumes. A classifier learns local landmark appearance, and Hough regressors are trained to aggregate neighborhood information to a precise landmark coordinate position. A non-parametric geometric model encodes the spatial relationships between the landmarks and derives a topology which connects mutually predictive landmarks. During the global search we classify all voxels in the query volume, and perform regression-based agglomeration of landmark probabilities to highly accurate and specific candidate points at potential landmark locations. We encode the candidates’ weights together with the conformity of the connecting edges to the learnt geometric model in a Markov Random Field (MRF). By solving the corresponding discrete optimization problem, the most probable location for each model landmark is found in the query volume. We show that this approach is able to consistently localize the model landmarks despite the complex and repetitive character of the anatomical structures on three challenging data sets (hand radiographs, hand CTs, and whole body CTs), with a median localization error of 0.80 mm, 1.19 mm and 2.71 mm, respectively., Medical Image Analysis, 17 (8), ISSN:1361-8415, ISSN:1361-8423

Traumatic brain injury (TBI) can disrupt the white matter (WM) integrity in the brain, leading to functional and cognitive disruptions that may persist for years. There is considerable heterogeneity within the patient group, which complicates group analyses. Here we present improvements to a tract identification workflow, automated multi-atlas tract extraction (autoMATE), evaluating the effects of improved registration. Use of study-specific template improved group classification accuracy over the standard workflow. The addition of a multi-modal registration that includes information from diffusion weighted imaging (DWI), T1-weighted, and Fluid-Attenuated Inversion Recovery (FLAIR) further improved classification accuracy. We also examined whether particular tracts contribute more to group classification than others. Parts of the corpus callosum contributed most, and there were unexpected asymmetries between bilateral tracts.

We present an unsupervised three-dimensional feature clustering algorithm to gather the mTOP2016 challenge data into 3 groups. We use the brain MR-T1, diffusion tensor fractional anisotropy, and diffusion tensor mean diffusivity images provided by the mTOP2016 competition. A distance-based size constraint method for data clustering is used. The proposed approach achieves 0.267 adjusted rand index and 0.3556 homogeneity score within the 15 labeled subjects, corresponding to 10 correctly classified data items. Based on visual exploration of the data, we believe that a localized analysis of the lesion regions, using the computed tractography data, is a promising direction to pursue.

The relationship between spatially distributed fMRI patterns and experimental stimuli or tasks offers insights into cognitive processes beyond those traceable from individual local activations. The multivariate properties of the fMRI signals allow us to infer interactions among individual regions and to detect distributed activations of multiple areas. Detection of task-specific multivariate activity in fMRI data is an important open problem that has drawn much interest recently. In this paper, we study and demonstrate the benefits of random forest classifiers and the associated Gini importance measure for selecting voxel subsets that form a multivariate neural response. The Gini importance measure quantifies the predictive power of a particular feature when considered as part of the entire pattern. The measure is based on a random sampling of fMRI time points and voxels. As a consequence the resulting voxel score, or Gini contrast, is highly reproducible and reliably includes all informative features. The method does not rely on a priori assumptions about the signal distribution, a specific statistical or functional model or regularization. Instead, it uses the predictive power of features to characterize their relevance for encoding task information. The Gini contrast offers an additional advantage of directly quantifying the task-relevant information in a multiclass setting, rather than reducing the problem to several binary classification subproblems. In a multicategory visual fMRI study, the proposed method identified informative regions not detected by the univariate criteria, such as the t-test or the F-test. Including these additional regions in the feature set improves the accuracy of multicategory classification. Moreover, we demonstrate higher classification accuracy and stability of the detected spatial patterns across runs than the traditional methods such as the recursive feature elimination used in conjunction with support vector machines., National Science Foundation (U.S.) (Grant IIS/CRCNS 0904625), National Science Foundation (U.S.) (CAREER Grant 0642971), National Institutes of Health (U.S.) (NCRR NAC P41-RR13218), National Institutes of Health (U.S.) (Grant NIBIB NAMIC U54-EB005149), German Academy of Sciences Leopoldina (Fellowship Programme LPDS 2009-10)

Spatial priors, such as probabilistic atlases, play an important role in MRI segmentation. However, the availability of comprehensive, reliable and suitable manual segmentations for atlas construction is limited. We therefore propose a method for joint segmentation of corresponding regions of interest in a collection of aligned images that does not require labeled training data. Instead, a latent atlas, initialized by at most a single manual segmentation, is inferred from the evolving segmentations of the ensemble. The algorithm is based on probabilistic principles but is solved using partial differential equations (PDEs) and energy minimization criteria. We evaluate the method on two datasets, segmenting subcortical and cortical structures in a multi-subject study and extracting brain tumors in a single-subject multi-modal longitudinal experiment. We compare the segmentation results to manual segmentations, when those exist, and to the results of a state-of-the-art atlas-based segmentation method. The quality of the results supports the latent atlas as a promising alternative when existing atlases are not compatible with the images to be segmented., National Institutes of Health (U.S.) (National Institute for Biomedical Imaging and Bioengineering (U.S.)/National Alliance for Medical Image Computing (U.S.) U54-EB005149), National Institutes of Health (U.S.) (National Center for Research Resources (U.S.)/Neuroimaging Analysis Center (U.S.) P41-RR13218), National Institutes of Health (U.S.) (National Institute of Neurological Disorders and Stroke (U.S.) R01-NS051826), National Institutes of Health (U.S.) (National Center for Research Resources (U.S.)/Biomedical Informatics Research Network U24-RR021382), National Science Foundation (U.S.) (CAREER Award 0642971), German Academy of Sciences Leopoldina (Fellowship LPDS 2009-10), Academy of Finland (Grant 133611)