Are Artificial Intelligence-Assisted Three-Dimensional Histological Reconstructions Reliable for the Assessment of Trabecular Microarchitecture?

Abstract

Objectives: This study aimed to create a three-dimensional histological reconstruction through the AI-assisted classification of tissues and the alignment of serial sections. The secondary aim was to evaluate if the novel technique for histological reconstruction accurately replicated the trabecular microarchitecture of bone. This was performed by conducting micromorphometric measurements on the reconstruction and comparing the results obtained with those of microCT reconstructions. Methods: A bone biopsy sample was harvested upon re-entry following sinus floor augmentation. Following microCT scanning and histological processing, a modified version of the U-Net architecture was trained to categorize tissues on the sections. Detector-free local feature matching with transformers was used to create the histological reconstruction. The micromorphometric parameters were calculated using Bruker's CTAn software (version 1.18.8.0, Bruker, Kontich, Belgium) for both histological and microCT datasets. Results: Correlation coefficients calculated between the micromorphometric parameters measured on the microCT and histological reconstruction suggest a strong linear relationship between the two with p-values of 0.777, 0.717, 0.705, 0.666, and 0.687 for BV/TV, BS/TV, Tb.Pf Tb.Th, and Tb.Sp, respectively. Bland-Altman and mountain plots suggest good agreement between BV/TV measurements on the two reconstruction methods. Conclusions: This novel method for three-dimensional histological reconstruction provides researchers with a tool that enables the assessment of accurate trabecular microarchitecture and histological information simultaneously.

Keywords: artificial intelligence (AI); bone augmentation; dental implant; histomorphometry; microCT; three-dimensional histological reconstruction.

Figure 1

Example of the matching process within the final alignment. The number of matching keypoints for each in-silico microCT slice is represented by a blue dot. Notice that the slice that corresponds to the histological slide used in this example has an order of magnitude more matching keypoints then the rest of the microCT slices. This peak in the keypoint curve was approximated using a bell curve, represented by the orange line, the mean of which corresponds to the in-silico microCT slice matching the histology slice.

Figure 2

Visualization of the matched LoFTR keypoints between the two imaging modalities: an in silico microCT slice and the histology; this specific match corresponds to the peak of the curve presented in Figure 1. Notice that there are only parallel lines connecting the keypoints; crossing lines would indicate false matches.

Figure 3

U-Net architecture used in tissue segmentation. Since each 2 × 2 MaxPooling layer halves the resolution of the feature map, the resolution of the bottleneck layer is 32 × 32 pixels. The blue boxes correspond to multi-channel feature maps with the number of channels noted either above or below the boxes. The white boxes represent the concatenation between the upsampled (Conv2DTransposed) feature map from the expanding part and the copied feature map from the contracting path.

Figure 4

Side-by-side examples showing the inputs and outputs for the tissue segmentation model. The first column, named “Original”, depicts the tile from the histology, overlayed with the handmade annotations (bone outlined with red and bone graft material outlined with green). The second column, named “Mask”, shows the annotations as segmentation masks. These masks delineate three histological categories—bone tissue (highlighted in red), bone graft material (in green), and non-mineralized tissue (in blue). The third column, named “Mask Predicted”, depicts examples of the U-Net model’s outputs, with paler colors indicating smaller probabilities for a given class. The fourth column, named “Mask Predicted smoothed & thresholded”, depicts the final annotation following the post-processing consisting of bilateral filtering (”smoothing”) and thresholding, which makes each pixel belong to the single most probable class only.

Figure 5

Bland-Altman (BA) plots on the left and mountain plots on the right comparing the microCT tiles and the 3D histological reconstruction tiles for the measured micromorphometrical parameters percent bone volume, bone surface/bone volume ratio, trabecular pattern factor, trabecular thickness, and trabecular separation. On the BA plots, the mean difference is indicated by a solid line, the 95% limits of agreements are indicated by dashed lines, and the fitted linear regression is indicated by a dotted line along with its 95% confidence band. Each parameter has close to zero bias (the slope of the mean–difference curve), but further statistical tests showed that this hypothesis cannot be rejected only for the percent bone volume; for the other parameters, the difference from zero bias is significant. On the mountain plots, the median difference is marked with a solid line and the 95% confidence interval is marked with dashed lines. The median difference indicates the bias between the two imaging methods; the closer it is to zero, the smaller the bias. Only the trabecular thickness measurements show a larger bias; however, all distributions show fairly long tails.