X-ray multiscale 3D neuroimaging to quantify cellular aging and neurodegeneration postmortem in a model of Alzheimer's disease.

PURPOSE: Modern neuroimaging lacks the tools necessary for whole-brain, anatomically dense neuronal damage screening. An ideal approach would include unbiased histopathologic identification of aging and neurodegenerative disease. METHODS: We report the postmortem application of multiscale X-ray phase-contrast computed tomography (X-PCI-CT) for the label-free and dissection-free organ-level to intracellular-level 3D visualization of distinct single neurons and glia. In deep neuronal populations in the brain of aged wild-type and of 3xTgAD mice (a triply-transgenic model of Alzheimer's disease), we quantified intracellular hyperdensity, a manifestation of aging or neurodegeneration. RESULTS: In 3xTgAD mice, the observed hyperdensity was identified as amyloid-ß and hyper-phosphorylated tau protein deposits with calcium and iron involvement, by correlating the X-PCI-CT data to immunohistochemistry, X-ray fluorescence microscopy, high-field MRI, and TEM. As a proof-of-concept, X-PCI-CT was used to analyze hippocampal and cortical brain regions of 3xTgAD mice treated with LY379268, selective agonist of group II metabotropic glutamate receptors (mGlu2/3 receptors). Chronic pharmacologic activation of mGlu2/3 receptors significantly reduced the hyperdensity particle load in the ventral cortical regions of 3xTgAD mice, suggesting a neuroprotective effect with locoregional efficacy. CONCLUSIONS: This multiscale micro-to-nano 3D imaging method based on X-PCI-CT enabled identification and quantification of cellular and sub-cellular aging and neurodegeneration in deep neuronal and glial cell populations in a transgenic model of Alzheimer's disease. This approach quantified the localized and intracellular neuroprotective effects of pharmacological activation of mGlu2/3 receptors.

High-Spatial-Resolution Three-dimensional Imaging of Human Spinal Cord and Column Anatomy with Postmortem X-ray Phase-Contrast Micro-CT.

Background Modern high-spatial-resolution radiologic methods enable increasingly detailed volumetric postmortem investigations of human neuroanatomy for diagnostic, research, and educational purposes. Purpose To evaluate the viability of postmortem x-ray phase-contrast micro-CT to provide tissue-conserving, high-spatial-resolution, three-dimensional neuroimaging of the human spinal cord and column by comparing quality of x-ray phase-contrast micro-CT images of nondissected Thiel-embalmed human spines with images of extracted formalin-fixed human spinal cords. Specific focus was placed on assessing the detection of micrometric spinal cord soft-tissue structure and vasculature. Materials and Methods In this study from August 2015 to August 2019, three Thiel-embalmed human spinal column samples, unilaterally perfused with an iodinated vascular contrast agent, and three extracted formalin-fixed spinal cord samples were imaged postmortem at a synchrotron radiation facility. Propagation-based x-ray phase-contrast micro-CT was used with monochromatic 60-keV x-rays and a detector with either 46-µm or 8-µm pixel sizes. A single-distance phase-retrieval algorithm was applied to the acquired CT projection images in advance of filtered back projection CT reconstruction. The influence on image quality of Thiel versus formalin embalming was examined, and images were qualitatively evaluated in terms of the value of their anatomic representations. Results The x-ray phase-contrast micro-CT of Thiel-embalmed samples resulted in soft-tissue contrast within the vertebral canal, despite evident nervous tissue deterioration after Thiel embalming. Gross spinal cord anatomy, spinal meninges, contrast agent-enhanced spinal vasculature, and spinal nerves were all well rendered alongside surrounding vertebral bone structure. The x-ray phase-contrast micro-CT of formalin-fixed boneless cords led to much higher gray versus white matter contrast and to microscale visualization of deep medullary vasculature and neuronal perikarya. Conclusion This work demonstrated the use of x-ray phase-contrast micro-CT for detailed volumetric anatomic visualization of embalmed human spines. The method provided three-dimensional display of bone, nervous tissue, and vasculature at microscale resolutions without exogenous contrast agents. © RSNA, 2020 Online supplemental material is available for this article.

Multiscale X-ray phase contrast imaging of human cartilage for investigating osteoarthritis formation.

BACKGROUND: The evolution of cartilage degeneration is still not fully understood, partly due to its thinness, low radio-opacity and therefore lack of adequately resolving imaging techniques. X-ray phase-contrast imaging (X-PCI) offers increased sensitivity with respect to standard radiography and CT allowing an enhanced visibility of adjoining, low density structures with an almost histological image resolution. This study examined the feasibility of X-PCI for high-resolution (sub-) micrometer analysis of different stages in tissue degeneration of human cartilage samples and compare it to histology and transmission electron microscopy. METHODS: Ten 10%-formalin preserved healthy and moderately degenerated osteochondral samples, post-mortem extracted from human knee joints, were examined using four different X-PCI tomographic set-ups using synchrotron radiation the European Synchrotron Radiation Facility (France) and the Swiss Light Source (Switzerland). Volumetric datasets were acquired with voxel sizes between 0.7 × 0.7 × 0.7 and 0.1 × 0.1 × 0.1 µm3. Data were reconstructed by a filtered back-projection algorithm, post-processed by ImageJ, the WEKA machine learning pixel classification tool and VGStudio max. For correlation, osteochondral samples were processed for histology and transmission electron microscopy. RESULTS: X-PCI provides a three-dimensional visualization of healthy and moderately degenerated cartilage samples down to a (sub-)cellular level with good correlation to histologic and transmission electron microscopy images. X-PCI is able to resolve the three layers and the architectural organization of cartilage including changes in chondrocyte cell morphology, chondrocyte subgroup distribution and (re-)organization as well as its subtle matrix structures. CONCLUSIONS: X-PCI captures comprehensive cartilage tissue transformation in its environment and might serve as a tissue-preserving, staining-free and volumetric virtual histology tool for examining and chronicling cartilage behavior in basic research/laboratory experiments of cartilage disease evolution.

Multiscale pink-beam microCT imaging at the ESRF-ID17 biomedical beamline.

Recent trends in hard X-ray micro-computed tomography (microCT) aim at increasing both spatial and temporal resolutions. These challenges require intense photon beams. Filtered synchrotron radiation beams, also referred to as `pink beams', which are emitted by wigglers or bending magnets, meet this need, owing to their broad energy range. In this work, the new microCT station installed at the biomedical beamline ID17 of the European Synchrotron is described and an overview of the preliminary results obtained for different biomedical-imaging applications is given. This new instrument expands the capabilities of the beamline towards sub-micrometre voxel size scale and simultaneous multi-resolution imaging. The current setup allows the acquisition of tomographic datasets more than one order of magnitude faster than with a monochromatic beam configuration.

Quantitative Assessment of Degenerative Cartilage and Subchondral Bony Lesions in a Preserved Cadaveric Knee: Propagation-Based Phase-Contrast CT Versus Conventional MRI and CT.

OBJECTIVE: The aim of this study was to quantitatively assess hyaline cartilage and subchondral bone conditions in a fully preserved cadaveric human knee joint using high-resolution x-ray propagation-based phase-contrast imaging (PBI) CT and to compare the performance of the new technique with conventional CT and MRI. MATERIALS AND METHODS: A cadaveric human knee was examined using an x-ray beam of 60 keV, a detector with a 90-mm2 FOV, and a pixel size of 46 × 46 µm2. PBI CT images were reconstructed with both the filtered back projection algorithm and the equally sloped tomography method. Conventional 3-T MRI and CT were also performed. Measurements of cartilage thickness, cartilage lesions, International Cartilage Repair Society scoring, and detection of subchondral bone changes were evaluated. Visual inspection of the specimen akin to arthroscopy was conducted and served as a standard of reference for lesion detection. RESULTS: Loss of cartilage height was visible on PBI CT and MRI. Quantification of cartilage thickness showed a strong correlation between the two modalities. Cartilage lesions appeared darker than the adjacent cartilage on PBI CT. PBI CT showed similar agreement to MRI for depicting cartilage substance defects or lesions compared with the visual inspection. The assessment of subchondral bone cysts showed moderate to strong agreement between PBI CT and CT. CONCLUSION: In contrast to the standard clinical methods of MRI and CT, PBI CT is able to simultaneously depict cartilage and bony changes at high resolution. Though still an experimental technique, PBI CT is a promising high-resolution imaging method to evaluate comprehensive changes of osteoarthritic disease in a clinical setting.

Characterization of a sCMOS-based high-resolution imaging system.

The detection system is a key part of any imaging station. Here the performance of the novel sCMOS-based detection system installed at the ID17 biomedical beamline of the European Synchrotron Radiation Facility and dedicated to high-resolution computed-tomography imaging is analysed. The system consists of an X-ray-visible-light converter, a visible-light optics and a PCO.Edge5.5 sCMOS detector. Measurements of the optical characteristics, the linearity of the system, the detection lag, the modulation transfer function, the normalized power spectrum, the detective quantum efficiency and the photon transfer curve are presented and discussed. The study was carried out at two different X-ray energies (35 and 50 keV) using both 2× and 1× optical magnification systems. The final pixel size resulted in 3.1 and 6.2 µm, respectively. The measured characteristic parameters of the PCO.Edge5.5 are in good agreement with the manufacturer specifications. Fast imaging can be achieved using this detection system, but at the price of unavoidable losses in terms of image quality. The way in which the X-ray beam inhomogeneity limited some of the performances of the system is also discussed.

Single-image phase retrieval using an edge illumination X-ray phase-contrast imaging setup.

A method is proposed which enables the retrieval of the thickness or of the projected electron density of a sample from a single input image acquired with an edge illumination phase-contrast imaging setup. The method assumes the case of a quasi-homogeneous sample, i.e. a sample with a constant ratio between the real and imaginary parts of its complex refractive index. Compared with current methods based on combining two edge illumination images acquired in different configurations of the setup, this new approach presents advantages in terms of simplicity of acquisition procedure and shorter data collection time, which are very important especially for applications such as computed tomography and dynamical imaging. Furthermore, the fact that phase information is directly extracted, instead of its derivative, can enable a simpler image interpretation and be beneficial for subsequent processing such as segmentation. The method is first theoretically derived and its conditions of applicability defined. Quantitative accuracy in the case of homogeneous objects as well as enhanced image quality for the imaging of complex biological samples are demonstrated through experiments at two synchrotron radiation facilities. The large range of applicability, the robustness against noise and the need for only one input image suggest a high potential for investigations in various research subjects.

Boundary value problem for phase retrieval from unidirectional X-ray differential phase images.

The phase retrieval problem can be reduced to the second order partial differential equation. In order to retrieve the absolute values of the X-ray phase and to minimize the reconstruction artifacts we defined the mixed inhomogeneous boundary condition using available a priori information about the sample. Finite element technique was used to solve the boundary value problem. The approach is validated on numerical and experimental phantoms. In order to demonstrate a possible application of the method, we have processed an entire tomographic set of differential phase images and estimated the magnitude of the refractive index decrement for some tissues inside complex biomedical samples.

High-resolution, low-dose phase contrast X-ray tomography for 3D diagnosis of human breast cancers.

Mammography is the primary imaging tool for screening and diagnosis of human breast cancers, but ~10-20% of palpable tumors are not detectable on mammograms and only about 40% of biopsied lesions are malignant. Here we report a high-resolution, low-dose phase contrast X-ray tomographic method for 3D diagnosis of human breast cancers. By combining phase contrast X-ray imaging with an image reconstruction method known as equally sloped tomography, we imaged a human breast in three dimensions and identified a malignant cancer with a pixel size of 92 µm and a radiation dose less than that of dual-view mammography. According to a blind evaluation by five independent radiologists, our method can reduce the radiation dose and acquisition time by ~74% relative to conventional phase contrast X-ray tomography, while maintaining high image resolution and image contrast. These results demonstrate that high-resolution 3D diagnostic imaging of human breast cancers can, in principle, be performed at clinical compatible doses.

Tomographic reconstruction of the refractive index with hard X-rays: an efficient method based on the gradient vector-field approach.

The refractive-index gradient vector field approach establishes a connection between a tomographic data set of differential phase contrast images and the distribution of the partial spatial derivatives of the refractive index in an object. The reconstruction of the refractive index in a plane requires the integration of its gradient field. This work shows how this integration can be efficiently performed by converting the problem to the Poisson equation, which can be accurately solved even in the case of noisy and large datasets. The performance of the suggested method is discussed and demonstrated experimentally by computing the refractive index distribution in both a simple plastic phantom and a complex biological sample. The quality of the reconstruction is evaluated through the direct comparison with other commonly used methods. To this end, the refractive index is retrieved from the same data set using also (1) the filtered backprojection algorithm for gradient projections, and (2) the regularized phase-retrieval procedure. Results show that the gradient vector field approach combined with the developed integration technique provides a very accurate depiction of the sample internal structure. Contrary to the two other techniques, the considered method does not require a preliminary phase-retrieval and can be implemented with any advanced computer tomography algorithm. In this work, analyzer-based phase contrast images are used for demonstration. Results, however, are generally valid and can be applied for processing differential phase-contrast tomographic data sets obtained with other phase-contrast imaging techniques.

Computer-aided diagnosis for phase-contrast X-ray computed tomography: quantitative characterization of human patellar cartilage with high-dimensional geometric features.

Phase-contrast computed tomography (PCI-CT) has shown tremendous potential as an imaging modality for visualizing human cartilage with high spatial resolution. Previous studies have demonstrated the ability of PCI-CT to visualize (1) structural details of the human patellar cartilage matrix and (2) changes to chondrocyte organization induced by osteoarthritis. This study investigates the use of high-dimensional geometric features in characterizing such chondrocyte patterns in the presence or absence of osteoarthritic damage. Geometrical features derived from the scaling index method (SIM) and statistical features derived from gray-level co-occurrence matrices were extracted from 842 regions of interest (ROI) annotated on PCI-CT images of ex vivo human patellar cartilage specimens. These features were subsequently used in a machine learning task with support vector regression to classify ROIs as healthy or osteoarthritic; classification performance was evaluated using the area under the receiver-operating characteristic curve (AUC). SIM-derived geometrical features exhibited the best classification performance (AUC, 0.95 ± 0.06) and were most robust to changes in ROI size. These results suggest that such geometrical features can provide a detailed characterization of the chondrocyte organization in the cartilage matrix in an automated and non-subjective manner, while also enabling classification of cartilage as healthy or osteoarthritic with high accuracy. Such features could potentially serve as imaging markers for evaluating osteoarthritis progression and its response to different therapeutic intervention strategies.

An efficient numerical tool for dose deposition prediction applied to synchrotron medical imaging and radiation therapy.

Medical imaging and radiation therapy are widely used synchrotron-based techniques which have one thing in common: a significant dose delivery to typically biological samples. Among the ways to provide the experimenters with image guidance techniques indicating optimization strategies, Monte Carlo simulation has become the gold standard for accurately predicting radiation dose levels under specific irradiation conditions. A highly important hampering factor of this method is, however, its slow statistical convergence. A track length estimator (TLE) module has been coded and implemented for the first time in the open-source Monte Carlo code GATE/Geant4. Results obtained with the module and the procedures used to validate them are presented. A database of energy-absorption coefficients was also generated, which is used by the TLE calculations and is now also included in GATE/Geant4. The validation was carried out by comparing the TLE-simulated doses with experimental data in a synchrotron radiation computed tomography experiment. The TLE technique shows good agreement versus both experimental measurements and the results of a classical Monte Carlo simulation. Compared with the latter, it is possible to reach a pre-defined statistical uncertainty in about two to three orders of magnitude less time for complex geometries without loss of accuracy.

X-ray phase contrast imaging of objects with subpixel-size inhomogeneities: a geometrical optics model.

Several x-ray phase contrast extraction algorithms use a set of images acquired along the rocking curve of a perfect flat analyzer crystal to study the internal structure of objects. By measuring the angular shift of the rocking curve peak, one can determine the local deflections of the x-ray beam propagated through a sample. Additionally, some objects determine a broadening of the crystal rocking curve, which can be explained in terms of multiple refraction of x rays by many subpixel-size inhomogeneities contained in the sample. This fact may allow us to differentiate between materials and features characterized by different refraction properties. In the present work we derive an expression for the beam broadening in the form of a linear integral of the quantity related to statistical properties of the dielectric susceptibility distribution function of the object.

X-Ray Phase Contrast 3D Virtual Histology: Evaluation of Lung Alterations After Microbeam Irradiation.

PURPOSE: This study provides the first experimental application of multiscale 3-dimensional (3D) x-ray phase contrast imaging computed tomography (XPCI-CT) virtual histology for the inspection and quantitative assessment of the late-stage effects of radio-induced lesions on lungs in a small animal model. METHODS AND MATERIALS: Healthy male Fischer rats were irradiated with x-ray standard broad beams and microbeam radiation therapy, a high-dose rate (14 kGy/s), FLASH spatially fractionated x-ray therapy to avoid beamlet smearing owing to cardiosynchronous movements of the organs during the irradiation. After organ dissection, ex vivo XPCI-CT was applied to all the samples and the results were quantitatively analyzed and correlated to histologic data. RESULTS: XPCI-CT enables the 3D visualization of lung tissues with unprecedented contrast and sensitivity, allowing alveoli, vessel, and bronchi hierarchical visualization. XPCI-CT discriminates in 3D radio-induced lesions such as fibrotic scars and Ca/Fe deposits and allows full-organ accurate quantification of the fibrotic tissue within the irradiated organs. The radiation-induced fibrotic tissue content is less than 10% of the analyzed volume for all microbeam radiation therapy-treated organs and reaches 34% in the case of irradiations with 50 Gy using a broad beam. CONCLUSIONS: XPCI-CT is an effective imaging technique able to provide detailed 3D information for the assessment of lung pathology and treatment efficacy in a small animal model.

Imaging of the angular-dependent coherent-scatter cross section with analyzer crystal: a Monte Carlo simulation.

A nontomography approach for the measurement of angular-dependent coherent-scatter cross section of x rays (E≃40-80 keV) is described. It is shown that an analyzer crystal, which is proposed to be used for the sampling of the cross section, simultaneously provides information about the location of the scattering volume inside the object. A numerical simulation demonstrates that this method can be applied for nondestructive analysis of an object's internal structure.

A Multi-Scale and Multi-Technique Approach for the Characterization of the Effects of Spatially Fractionated X-ray Radiation Therapies in a Preclinical Model.

The purpose of this study is to use a multi-technique approach to detect the effects of spatially fractionated X-ray Microbeam (MRT) and Minibeam Radiation Therapy (MB) and to compare them to seamless Broad Beam (BB) irradiation. Healthy- and Glioblastoma (GBM)-bearing male Fischer rats were irradiated in-vivo on the right brain hemisphere with MRT, MB and BB delivering three different doses for each irradiation geometry. Brains were analyzed post mortem by multi-scale X-ray Phase Contrast Imaging-Computed Tomography (XPCI-CT), histology, immunohistochemistry, X-ray Fluorescence (XRF), Small- and Wide-Angle X-ray Scattering (SAXS/WAXS). XPCI-CT discriminates with high sensitivity the effects of MRT, MB and BB irradiations on both healthy and GBM-bearing brains producing a first-time 3D visualization and morphological analysis of the radio-induced lesions, MRT and MB induced tissue ablations, the presence of hyperdense deposits within specific areas of the brain and tumor evolution or regression with respect to the evaluation made few days post-irradiation with an in-vivo magnetic resonance imaging session. Histology, immunohistochemistry, SAXS/WAXS and XRF allowed identification and classification of these deposits as hydroxyapatite crystals with the coexistence of Ca, P and Fe mineralization, and the multi-technique approach enabled the realization, for the first time, of the map of the differential radiosensitivity of the different brain areas treated with MRT and MB. 3D XPCI-CT datasets enabled also the quantification of tumor volumes and Ca/Fe deposits and their full-organ visualization. The multi-scale and multi-technique approach enabled a detailed visualization and classification in 3D of the radio-induced effects on brain tissues bringing new essential information towards the clinical implementation of the MRT and MB radiation therapy techniques.

Absorption, refraction and scattering in analyzer-based imaging: comparison of different algorithms.

Many mathematical methods have been so far proposed in order to separate absorption, refraction and ultra-small angle scattering information in phase-contrast analyzer-based images. These algorithms all combine a given number of images acquired at different positions of the crystal analyzer along its rocking curve. In this paper a comprehensive quantitative comparison between five of the most widely used phase extraction algorithms based on the geometrical optics approximation is presented: the diffraction-enhanced imaging (DEI), the extended diffraction-enhanced imaging (E-DEI), the generalized diffraction-enhanced (G-DEI), the multiple-image radiography (MIR) and the Gaussian curve fitting (GCF). The algorithms are theoretically analyzed in terms of their validity conditions and experimentally compared by using geometrical phantoms providing various amounts of absorption, refraction and scattering. The presented work shows that, due to their specific validity conditions, the considered algorithms produce results that may greatly differ, especially in the case of highly refracting and/or highly scattering materials. The various extraction algorithms are also applied to images of a human bone-cartilage sample. The aim is to validate the results obtained on geometrical phantoms and prove the efficiency of the different algorithms for applications on biological samples.

Convolutional neuronal networks combined with X-ray phase-contrast imaging for a fast and observer-independent discrimination of cartilage and liver diseases stages.

We applied transfer learning using Convolutional Neuronal Networks to high resolution X-ray phase contrast computed tomography datasets and tested the potential of the systems to accurately classify Computed Tomography images of different stages of two diseases, i.e. osteoarthritis and liver fibrosis. The purpose is to identify a time-effective and observer-independent methodology to identify pathological conditions. Propagation-based X-ray phase contrast imaging WAS used with polychromatic X-rays to obtain a 3D visualization of 4 human cartilage plugs and 6 rat liver samples with a voxel size of 0.7 × 0.7 × 0.7 µm3 and 2.2 × 2.2 × 2.2 µm3, respectively. Images with a size of 224 × 224 pixels are used to train three pre-trained convolutional neuronal networks for data classification, which are the VGG16, the Inception V3, and the Xception networks. We evaluated the performance of the three systems in terms of classification accuracy and studied the effect of the variation of the number of inputs, training images and of iterations. The VGG16 network provides the highest classification accuracy when the training and the validation-test of the network are performed using data from the same samples for both the cartilage (99.8%) and the liver (95.5%) datasets. The Inception V3 and Xception networks achieve an accuracy of 84.7% (43.1%) and of 72.6% (53.7%), respectively, for the cartilage (liver) images. By using data from different samples for the training and validation-test processes, the Xception network provided the highest test accuracy for the cartilage dataset (75.7%), while for the liver dataset the VGG16 network gave the best results (75.4%). By using convolutional neuronal networks we show that it is possible to classify large datasets of biomedical images in less than 25 min on a 8 CPU processor machine providing a precise, robust, fast and observer-independent method for the discrimination/classification of different stages of osteoarthritis and liver diseases.

Establishing sample-preparation protocols for X-ray phase-contrast CT of rodent spinal cords: Aldehyde fixations and osmium impregnation.

BACKGROUND: Dense and unbiased cellular-resolution representations of extended volumetric central nervous system soft-tissue anatomy are difficult to obtain, even in experimental post-mortem settings. Interestingly, X-ray phase-contrast computed tomography (X-PCI-CT), an emerging soft-tissue-sensitive volumetric imaging technique, can provide multiscale organ- to cellular-level morphological visualizations of neuroanatomical structure. NEW METHOD: Here, we tested different nervous-tissue fixation procedures, conventionally used for transmission electron microscopy, to better establish X-PCI-CT-specific sample-preparation protocols. Extracted rat spinal medullas were alternatively fixed with a standard paraformaldehyde-only aldehyde-based protocol, or in combination with glutaraldehyde. Some specimens were additionally post-fixed with osmium tetroxide. Multiscale X-PCI-CT datasets were collected at several synchrotron radiation facilities, using state-of-the-art setups with effective image voxel sizes of 3.03 to 0.33 µm3, and compared to high-field magnetic resonance imaging, histology and vascular fluorescence microscopy data. RESULTS: Multiscale X-PCI-CT of aldehyde-fixed spinal cord specimens resulted in dense histology-like volumetric representations and quantifications of extended deep spinal micro-vascular networks and of intra-medullary cell populations. Osmium post-fixation increased intra-medullary contrast between white and gray-matter tissues, and enhanced delineation of intra-medullary cellular structure, e.g. axon fibers and motor neuron perikarya. COMPARISON WITH EXISTING METHODS: Volumetric X-PCI-CT provides complementary contrast and higher spatial resolution compared to 9.4 T MRI. X-PCI-CT's advantage over planar histology is the volumetric nature of the cellular-level data obtained, using samples much larger than those fit for volumetric vascular fluorescence microscopy. CONCLUSIONS: Deliberately choosing (post-)fixation protocols tailored for optimal nervous-tissue structural preservation is of paramount importance in achieving effective and targeted neuroimaging via the X-PCI-CT technique.

Deep transfer learning for characterizing chondrocyte patterns in phase contrast X-Ray computed tomography images of the human patellar cartilage.

Phase contrast X-ray computed tomography (PCI-CT) has been demonstrated to be effective for visualization of the human cartilage matrix at micrometer resolution, thereby capturing osteoarthritis induced changes to chondrocyte organization. This study aims to systematically assess the efficacy of deep transfer learning methods for classifying between healthy and diseased tissue patterns. We extracted features from two different convolutional neural network architectures, CaffeNet and Inception-v3 for characterizing such patterns. These features were quantitatively evaluated in a classification task measured by the area (AUC) under the Receiver Operating Characteristic (ROC) curve as well as qualitative visualization through a dimension reduction approach t-Distributed Stochastic Neighbor Embedding (t-SNE). The best classification performance, for CaffeNet, was observed when using features from the last convolutional layer and the last fully connected layer (AUCs >0.91). Meanwhile, off-the-shelf features from Inception-v3 produced similar classification performance (AUC >0.95). Visualization of features from these layers further confirmed adequate characterization of chondrocyte patterns for reliably distinguishing between healthy and osteoarthritic tissue classes. Such techniques, can be potentially used for detecting the presence of osteoarthritis related changes in the human patellar cartilage.