Processing Techniques for Scanning Electron Microscopy Imaging of Giant Cells from Giant Cell Tumors of Bone

Abstract

Giant cell tumor (GCT) of bone is a common benign lesion that causes significant morbidity due to the failure of modern medical and surgical treatment. Surface ultra-structures of giant cells (GCs) may help in distinguishing aggressive tumors from indolent GC lesions. This study aimed to standardize scanning electron microscopic (SEM) imaging of GC from GCT of bone. Fresh GCT collected in Dulbecco's Modified Eagle Medium was washed to remove blood, homogenized, or treated with collagenase to isolate the GCs. Mechanically homogenized and collagenase-digested GCs were imaged on SEM after commonly used drying methodologies such as air-drying, tetramethylsilane (TMS)-drying, freeze-drying, and critical point-drying (CPD) for the optimization of sample processing. The collagenase-treated samples yielded a greater number of isolated GC and showed better surface morphology in comparison to mechanical homogenization. Air-drying was associated with marked cell shrinkage, and freeze-dried samples showed severe cell damage. TMS methodology partially preserved the cell contour and surface structures, although the cell shape was distorted. GC images with optimum surface morphology including membrane folding and microvesicular structures on the surface were observed only in collagenase-treated and critical point-dried samples. Collagenase digestion and critical point/TMS-drying should be performed for optimal SEM imaging of individual GCs.

Application of convolutional neural networks towards nuclei segmentation in localization-based super-resolution fluorescence microscopy images

“…The original UNet study accomplished this using a loss function that weights edge loss more heavily than interior, forcing the network to better learn instance boundaries [ 11 ]. Other techniques include teaching a secondary CNN to learn the borders between clustered cells or the cell edges [ 21 , 30 , 31 ], thresholding, morphological processing, active contours [ 3 ] and watershed [ 32 , 33 ]. These methods are not always trained as part of the network, and are commonly applied during post-processing on the instance predictions made following network training.…”