Deep learning for early detection of pathological changes in X-ray bone microstructures: case of osteoarthritis.

Texture features are designed to quantitatively evaluate patterns of spatial distribution of image pixels for purposes of image analysis and interpretation. Unexplained variations in the texture patterns often lead to misinterpretation and undesirable consequences in medical image analysis. In this paper we explore the ability of machine learning (ML) methods to design a radiology test of Osteoarthritis (OA) at early stage when the number of patients' cases is small. In our experiments we use high-resolution X-ray images of knees in patients which were identified with Kellgren-Lawrence scores progressing from 1. The existing ML methods have provided a limited diagnostic accuracy, whilst the proposed Group Method of Data Handling strategy of Deep Learning has significantly extended the diagnostic test. The comparative experiments demonstrate that the proposed framework using the Zernike-based texture features has significantly improved the diagnostic accuracy on average by 11%. This allows us to conclude that the designed model for early diagnostic of OA will provide more accurate radiology tests, although new study is required when a large number of patients' cases will be available.

BrainGazer--visual queries for neurobiology research.

Neurobiology investigates how anatomical and physiological relationships in the nervous system mediate behavior. Molecular genetic techniques, applied to species such as the common fruit fly Drosophila melanogaster, have proven to be an important tool in this research. Large databases of transgenic specimens are being built and need to be analyzed to establish models of neural information processing. In this paper we present an approach for the exploration and analysis of neural circuits based on such a database. We have designed and implemented BrainGazer, a system which integrates visualization techniques for volume data acquired through confocal microscopy as well as annotated anatomical structures with an intuitive approach for accessing the available information. We focus on the ability to visually query the data based on semantic as well as spatial relationships. Additionally, we present visualization techniques for the concurrent depiction of neurobiological volume data and geometric objects which aim to reduce visual clutter. The described system is the result of an ongoing interdisciplinary collaboration between neurobiologists and visualization researchers.

Fully Convolutional Architectures for Multiclass Segmentation in Chest Radiographs.

The success of deep convolutional neural networks (NNs) on image classification and recognition tasks has led to new applications in very diversified contexts, including the field of medical imaging. In this paper, we investigate and propose NN architectures for automated multiclass segmentation of anatomical organs in chest radiographs (CXRs), namely for lungs, clavicles, and heart. We address several open challenges including model overfitting, reducing number of parameters, and handling of severely imbalanced data in CXR by fusing recent concepts in convolutional networks and adapting them to the segmentation problem task in CXR. We demonstrate that our architecture combining delayed subsampling, exponential linear units, highly restrictive regularization, and a large number of high-resolution low-level abstract features outperforms state-of-the-art methods on all considered organs, as well as the human observer on lungs and heart. The models use a multiclass configuration with three target classes and are trained and tested on the publicly available Japanese Society of Radiological Technology database, consisting of 247 X-ray images the ground-truth masks for which are available in the segmentation in CXR database. Our best performing model, trained with the loss function based on the Dice coefficient, reached mean Jaccard overlap scores of 95% for lungs, 86.8% for clavicles, and 88.2% for heart. This architecture outperformed the human observer results for lungs and heart.

Automated landmarking and labeling of fully and partially scanned spinal columns in CT images.

The spinal column is one of the most distinguishable structures in CT scans of the superior part of the human body. It is not necessary to segment the spinal column in order to use it as a frame of reference. It is sufficient to place landmarks and the appropriate anatomical labels at intervertebral disks and vertebrae. In this paper, we present an automated system for landmarking and labeling spinal columns in 3D CT datasets. We designed this framework with two goals in mind. First, we relaxed input data requirements found in the literature, and we label both full and partial spine scans. Secondly, we intended to fulfill the performance requirement for daily clinical use and developed a high throughput system capable of processing thousands of slices in just a few minutes. To accomplish the aforementioned goals, we encoded structural knowledge from training data in probabilistic boosting trees and used it to detect efficiently the spinal canal, intervertebral disks, and three reference regions responsible for initializing the landmarking and labeling. Final landmarks and labels are selected by Markov Random Field-based matches of newly introduced 3-disk models. The framework has been tested on 36 CT images having at least one of the regions around the thoracic first ribs, the thoracic twelfth ribs, or the sacrum. In an average time of 2 min, we achieved a correct labeling in 35 cases with precision of 99.0% and recall of 97.2%. Additionally, we present results assuming none of the three reference regions could be detected.

Entropy-optimized texture models.

In order to robustly match a statistical model of shape and appearance (e.g. AAM) to an unseen image, it is crucial to employ a robust model fittness measure. Dense sampling of texture over the region covered by the shape of interest makes comparison of model and image in principle robust. However, when merely texture differences are taken into account, problems with low contrast regions, fuzzy features, global intensity variations, and irregularly varying structures occur. In this paper we introduce a novel entropy-optimized texture model (ETM). We map gray values of training images such that pixels represent anatomical structures optimally in terms of information entropy. We match the ETM to unseen images employing Bayes' law. We validate our approach using four training sets (hearts in basal region, hearts in mid region, brain ventricles, and lumbar vertebrae) and conclude that ETMs perform better than AAMs. Not only they reduce the average point-to-contour error, they are better suited to cope with large texture variances due to different scanners and even modalities.