Answers are grounded in this paper and cite their sources.
Over the last 20 years, the use of X-ray tomography (XRT) in life sciences research has steadily increased. Research interests as varied as biomechanics, cancer research, vascular research and plant science are now making use of this non-destructive imaging approach both as a stand-alone technique and as part of a multimodal imaging workflow [1]. For histology applications, the key qualities of an XRT scan are high signal-to-noise ratios and resolution, since accessing 3D views of tissue specimens for assessment or potential diagnostics requires precise scans that deliver the highest possible confidence in the results. Increasing performance and throughput of lab-based XRT can be achieved at multiple points in the workflow; the X-ray source, the detection mechanism or the way in which the 2D projections are reconstructed into a 3D volume (the computed tomography (CT)). Most commercial lab-based XRT systems are classical micro-CT instruments that rely on reducing the distance between the specimen and the X-ray source to increase resolution. The data is usually reconstructed using the traditional Feldkamp-Davis-Kress (FDK) algorithm [2] which generates good quality images but requires a high number of projections, and/or long exposure times per projection to reduce image noise and artifacts. In contrast, the X-ray microscope (XRM) has a unique optical concept that includes scintillator-coupled objectives lenses which provide a means to generate high resolution scans of the interior of the sample [3]. Use of the very latest scintillator-coupled 40x objective lens optimized for highest resolution in the lower energies is demonstrated here using a skin biopsy specimen. In addition to the optical magnification, recent advances in reconstruction technologies are also revealing new insights. Machine learning is now well established for image analysis challenges such as segmentation [4]. The use of a trained neural network for improving the precision of segmentation is powerful and saves vast amounts of time when compared to manual approaches that would otherwise be required. Here we demonstrate the use of a deep learning reconstruction method (DeepRecon) to significantly increase signal-to-noise ratio through the implementation of pretrained neural networks. We demonstrate that DeepRecon extracts signals from low-dose data more efficiently than the conventional FDK method, which is sensitive to noise and prone to aliasing image artifacts. Representative images from Deep Recon reconstructions of skin biopsies show clearer boundaries and reveal structures that were otherwise shrouded by noise. A comparison of these results with those obtained by running a smoothing filter on a single virtual slice of the FDK reconstructed data show the superior efficacy of noise reduction by the deep learning reconstruction approach [5]. Heavy metal stained, resin embedded mouse skin punch biopsy (2mm diameter) imaged non-destructively using the 40xP objective of the ZEISS Xradia Versa XRM. The same dataset was reconstructed either using traditional FDK reconstruction (a) or deep learning based DeepRecon reconstruction using a convolutional neural network model trained on the same dataset (c). A smoothing filter was applied to the FDK reconstructed dataset to compare results (b). Yellow boxes in the top row of images represent the magnified regions shown beneath, and intensity line profiles of the indicated red lines are shown at the bottom.
No takes yet. Share an insight, caveat, or question.
Manser et al. (2023) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: