Keywords

1 Introduction

The increase of drug-resistant strains of Mycobacterium Tuberculosis (Tb) claims for new effective antibiotic combinations. The selection of compound candidates is speeded up by the use of animal models that accurately reflect the pathological progression of pulmonary tuberculosis [1,2,3].

The hallmark of tuberculosis is the formation of organized aggregates of immune cells, known as granulomas. In the presence of stimuli, tuberculous granulomas are the host-protective structures formed to contain infection. These containments act as barriers preventing the penetration of chemotherapeutic eradication agents, and also as an incubator for bacillus proliferation.

In granulomas, bacteria are predominantly found evading immune defences: intracellularly, inside macrophages and neutrophils and extracellularly, in interstitial tissue. Necrotizing granulomas present a caseous centre which constitutes the reservoir from which large bacterial numbers emerge. Furthermore, Tb can also provoke an inflammatory response in lung tissue which is subjected to repair, a niche that is often ignored [4, 5].

Thus, drug efficacy depends on its ability to reach both their extracellular and intracellular targets, penetrating and permeating complex lung lesion compartments.

The classical bacteriological examination used for human subjects such as the tuberculin skin test or the sputum culture are not available for mouse models. Classically, histopathological tests are performed at significant points during the experiment to estimate the number of viable bacteria or fungal cells in the samples [1, 6,7,8,9,10,11,12].

In drug development studies, in vivo low-dose high-resolution micro-CT imaging allows to follow up the advance/recession of the infection in terms of disease tissue extent [13, 14], independently on the type of lesion. Well established biomarkers are based on the intensity thresholding between the healthy and diseased lung parenchyma in thoracic micro-CT scans [6,7,8,9, 11, 12, 15]. Among the variety of texture features, the ones based on the grey level co-occurrence matrix (GLCM) [16] are proved to be especially useful in our context [17,18,19,20]. The information provided by these biomarkers allows to follow up the host response over time per subject of the experimental protocol, reducing the need of the histopathological evaluation. The strength of longitudinal studies is the deep understanding of the disease mechanisms and the structural changes it causes in the damaged parenchymal tissue.

The nomenclature stablished by CT biomarkers defines the mid-high intensity lung regions as soft diseased tissue and the high intensity lung regions as hard diseased tissue, remarking that the higher the diseased volume, the higher the disease burden (soft and hard). More specifically, the larger the hard volume, the lower responsiveness to treatment [10]. However, the relationship between grey levels and cellular composition of the lesion is not yet defined. We believe that the prognosis information given by those known biomarkers reflects underlying histopathology of the disease. For this reason, in this work, we propose a thresholding protocol which translates the histopathological segmentation to the CT images for the classification of granuloma intensities by their cellular composition. This approach enables the detection and stratification of tuberculosis involvement in micro-CT volumes of excised mouse lungs and opens the door to the assessment of treatment efficacy per granuloma composition.

2 Materials

For this work, we used lungs excised from two females C57BL/6J mice using procedures approved by the Animal Experimentation Ethics Committee of Hospital General Universitario Gregorio Marañón, Madrid, Spain and performed according to EU directive 2010/63/EU and national regulations (RD 53/2013). One was inoculated with the virulent strain of Tb H37Rv at the age of ten weeks. Both subjects were sacrificed eight weeks after the intratracheal insult. At that time, the infected mouse reached the chronic phase of the disease.

The preparation for histology consists on the immersion of the whole organ in paraffin blocks. The tissue was processed for fixation, dehydration, and wax immersion treatment. An iodinated-based staining step was added to the cycle before the wax immersion to enhance the CT contrast of the embedded organs.

When embedded in paraffine and before histology slicing, the two lungs were screened by micro CT scan. A standard micro-CT subsystem of a SuperArgus scanner (Sedecal Molecular Imaging, Madrid) was used (settings: 68 kV, 420 uA and soft-tissue filtering). We selected a 0.5° step-and-shoot protocol covering 360° and a multi-frame rate of eight frames per gantry position at 20 frames per second. These acquisition parameters lead to a total volume data of 4.27 GB (2.84 MB per frame, 80 frames in total) and a total acquisition time of sixteen minutes. Data-sets were reconstructed using the filtered back-projection (FBP) algorithm and an isotropic voxel resolution of 44 um. The axial CT slices were acquired parallel to the microtome slicing plane.

Once the acquisitions were finished, we processed four infected histological glass slides (from the disease model lung) and one healthy histological glass slide (from the healthy organ) with haematoxylin and eosin (He) stain which adds the contrast for nuclei, cytoplasm and extracellular matrix. The five slides were digitalized using the Aperio CS2 image capture device (Leica Biosystems, Nussloch, Germany) in tiled multi-resolution format. This format is the standard for virtual slides. By default, it is composed by three images of different resolution and each one stored as a separate layer within the image file. For the files on this work, the first layer corresponds to the full resolution image (at 40x magnification and 0.251 microns-per-pixel), the second level to a 25% of the original slide, meaning a 4:1 ratio, and the last level corresponds to a preview image called thumbnail.

3 Methods

The main steps of the proposed algorithm are presented in Fig. 1. The 3D micro-CT volumes and the 2D mid resolution histology images are the input datasets for the registration, annotation and Hounsfield Unit (HU) thresholding steps which result in the cellular segmentation of the micro-CT volume. The dissected Tb-infected mouse lung served as training sample for the protocol in Fig. 1 and the dissected healthy mouse lung served as thresholding testing sample.

Fig. 1.
figure 1

Thresholding algorithm: The paraffine block with the embedded organ is scanned by CT and then the block is sliced and digitalized. The CT slices are registered with the annotated histology slides. HU-based segmentation is done using the histopathological labelling information.

Full size image

3.1 Histological Annotation

An He slide (Fig. 2) corresponding to the right upper lobe of the infected animal was used to create and train the classifier of the information concerning lung tissue: healthy alveolar tissue appears as light pink and diseased tissue as dark violet. Other structures such as the airways and tracheal walls appear as the diseased tissue stained with a dark violet colour.

Fig. 2.
figure 2

Lung tissue appearance in histological images. Detail of lesions and cells represented by the colour scale in haematoxylin and eosin stain. Scale bars represent 2 mm and 100 um. (Color figure online)

Full size image

The granulomas from the murine model under study are characterized by unorganized lesions composed of lymphocyte or neutrophil foci forming a more defined cup, and of diffuse inflammatory cells (foamy macrophages) isolated or surrounding those cups. Thus, the images were annotated using four labels corresponding to background, healthy parenchyma (HP), foamy macrophages (FM) and neutrophil foci (NF).

An expert histopathologist created and trained the classifier using the Trainable Weka Segmentation tool [21]. The same tool was then used for labelling the other regions of the slide (the lobes which were not use for training) and the remaining histological slides. The segmentation accuracy has already been demonstrated by its utilization in microscopy [22,23,24,25,26,27,28].

3.2 CT-Histology Registration

Owing that the same excised sample has been acquired by CT and by microscopy imaging of the histology preparations, datasets are affine sets even after manipulating the paraffin block. The possible deformations may involve rotations, translations, scaling, and/or shears, which leads us to use the affine registration. By preserving collinearity and ratios of distances, the histological slide is deformed using a gradient descent optimization until the mutual information between it and the reference image is maximized. The mutual information is the similarity measure commonly used for registration of multimodality images and it is the parameter considered for the assessment of the optimum registration.

By finding the micro-CT slice correspondence and the transformation matrix, we can compare the structures and lesion compartments in the 2D histology slide and in the 3D micro-CT slice.

3.3 HU Thresholding by Histopathology

A preliminary slice segmentation based on the CT volume histogram served as basis for the fine threshold determination using the histopathological annotations. This step consisted on the identification of three tuberculosis-related tissue classes by iso-data thresholding: healthy tissue, soft tissue and hard tissue. An extra class is added for the background identification. The segmentation was done according to the classes defined in literature [9, 11, 29] and its main purpose is to delimit the regions of interest to avoid the misclassification induced by noise and image artefacts.

The four CT classes applied to each slice were then registered with their corresponding four histology classes using the previously derived affine transformation (Sect. 3.2). By separating the CT slices into four masks each corresponding to a normal distribution per tissue.

The intersection points of their probability density functions determine the HU thresholds. Radiodensity ranges can then be applied to the 3D CT volumes of the paraffin blocks to stratify the pulmonary tissues.

All methods presented in this section were developed in Matlab (Matlab Inc., Natick, MA, USA).

4 Results

The four histological slides from tuberculosis-infected subjects were classified using WEKA and registered with their corresponding 3D micro-CT slices (Fig. 3). From the four micro-CT slices, the HU tissue masks were extracted and the resulting normal distributions per tissue are shown in Fig. 4. The points of intersections of their probability density function are the micro-CT thresholds. These thresholds were used for the segmentation of the 3D micro-CT volumes, comprising the remaining 75 slices of the Tb-infected mouse lung and the 80 slices of dissected healthy mouse lung.

Fig. 3.
figure 3

Registered images: (first row) micro-CT slice, (second row) histological annotated slide registered with the micro-CT and (third row) the composite result with colour code: Black, background; blue, healthy parenchyma; green, foamy macrophages and red neutrophil foci. This leads to HU tissue masks guided by histological annotations. Image size 256 × 256 px. (Color figure online)

Full size image
Fig. 4.
figure 4

Histogram per tissue of the CT-histology registered slices (bars) and their derived probability distribution function (curves) of the normal distribution. Labels from the histology annotation have been preserved: black, background; blue, healthy parenchyma; green, foamy macrophages and red neutrophil foci. (Color figure online)

Full size image

To evaluate and validate the proposed histopathological thresholding protocol, two metrics were defined: the visual assessment of the resulting segmented volumes for both the infected and the healthy samples, and their quantitative assessment by the Jaccard index, a well-known similarity index.

The diagnostic ground truth used for evaluation is the manual segmentation of the 3D CT volumes by an expert radiologist, who used the intensity segmentation approach described in Gordaliza et al. [29] for evaluating disease burden. Table 1 gives a summary of the tissue segmentation thresholds used to build the ground truth and those computed by the presented histology-based approach.

Table 1. Thresholds for segmentation of 3D micro-CT volumes (HU).
Full size table

The qualitative assessment is depicted in Fig. 5 in which the composition of the segmented masks uses the colour scheme shown in Fig. 4. The absence of disease tissue (in the form of inflammatory response or of granulomas) can be visually assessed on the healthy volume (Fig. 5A).

Fig. 5.
figure 5

Composite of the classified 3D micro CT based on the HU distribution of the three tuberculosis-associated tissue volumes in a, healthy organ and in b, an infected organ. Black, background; blue, healthy parenchyma; green, foamy macrophages and red neutrophil foci. (Color figure online)

Full size image

The presence of disease tissue within healthy parenchyma can be assessed on the infected volume (Fig. 5B). The trachea and the oesophagus are also segmented as diseased tissue, since the radiodensities of those structures lay in the same ranges as the macrophages or the neutrophils. This circumstance is also present on the histological annotations.

The similarity between the radiologist and the histology-based segmentation was measured using the Jaccard index, a metric that quantifies the size of the intersection divided by the union of the sets under comparison: the closer to the unit, the higher similarity. Table 2 shows the obtained Jaccard indexes for each type of segmented tissue (healthy parenchyma, foamy macrophages, neutrophil foci) and all segmented tissue.

Table 2. Jaccard indexes between histology-based segmentation and radiologist decision.
Full size table

5 Discussion

Our results confirm that the proposed histopathology on x ray CT provides a satisfactory estimation of the granuloma cellular structure by statistically modelling the HU distribution and by registering the CT with histological slides. The proposed methodology automatically assigns thresholds to 3D micro-CT for revealing the presence, extent and appearance of the immune responses to a tuberculosis challenge. Lobes and tuberculous granulomas were segmented by healthy parenchyma, foamy macrophages (predominant in inflammatory response) and neutrophil foci.

Histological annotations based in intensity have misidentifications owing that multiple tissues are coded with the same colour range. In our case, the characteristic cells composing the walls of collapsed alveoli, trachea and airways are automatically annotated as diseased tissue, both in the 3D micro CT volume and in the 2D histological slide. Spots in the digitalized histology slide due to dust in the lens were also detected by the classifier as diseased tissue. Furthermore, the staining concentration also interferes the classification, preventing the extraction of a global set of thresholds for CT slices or He slides. Approaches worth to explore are those using texture to classify the different cell components on the histological slides and those using histogram equalization strategies to further generalize the HU thresholds for Tb-related tissues.

The correspondence between standard histology and molecular imaging techniques can be estimated using registration techniques tailored for multimodal images. Most approaches rely on rigid registrations (i.e., affine transformation), for the initial slice correspondence between the 2D histology and 3D micro-CT [30, 31]. The intact tissue preservation and the digitalization quality are critical aspects for an optimum registration.

There are multiple image-based procedures which benefit from the 3D/2D registration of the whole organ image volume with histological slides. For example, the correspondence between MRI and histopathology in prostate cancer detection [32, 33]. In this type of cancer, it is generally difficult to differentiate the cancerous and non-cancerous regions directly on the preoperative in vivo MR images and the information that histopathological images offer ease the planning tasks.

It has also been demonstrated that is a good method to compare bone structure measurements performed on micro-CT images and to check that the diagnosis is correct [30, 31]. As far as we know, few results have been presented for soft tissue imaging tasks such as the pulmonary tissue [34,35,36].

It must also be considered that there is not exact correspondence between the histology image and a slice in the micro-CT. The voxel size of our 2D micro-CT slice is 40 um, meaning that the image is a flat cross-section which integrates all the structures within 40 um depth, whereas the width of the histology samples is 3 um. This may lead to misclassification of lesions or patterns that appear in the histology slide but not in the CT slice. One common event is the edge effect. Edges in histology have high intensity values and sometimes misclassified as neutrophil focus. However, those patterns are not expressed at the micro-CT scale. The limitations mentioned will be addressed in future studies by considering multiple disease models, strains and disease mechanisms, which will extend the robustness and accuracy of the granuloma stratification. We will focus on increasing the sample size of the training dataset with the goal of translating the approach to in-vivo longitudinal studies for a robust prediction of treatment outcome. With such studies, the efficacy of a compound to penetrate any type of lesion can be tested and therefore, the time to find plausible candidates for clinical trials can be reduced.

To conclude, we have proposed a fully automatic method for a virtual histopathological analysis based on the radiodensities of tuberculosis involvement on whole mouse lungs. The statistical model profits from the expert’s semiquantitative histopathological annotations. The method has the potential to define the correspondence between the grey-level intensity-based biomarkers and the pathological manifestations of infectious diseases involving any organ.