Automatic nuclei segmentation of label-free chromatin-sensitive partial wave spectroscopic microscopy using convolution neural network with transformer

Chromatin-sensitive Partial Wave Spectroscopic (csPWS) microscopy is a label-free spectroscopic optical nanosensing technique used to study the dynamic chromatin packing structure in the cell nucleus. Chromatin, composed of DNA wrapped around histones to form nucleosomes, organizes into supranucleosomal disordered chromatin chains, which assemble into packing domains (PDs). Dysregulation of these chromatin PDs is a common factor in the transcriptional phenotypic plasticity by which cancer cells change their state without changing their genes. csPWS microscopy quantifies the relationship between the genomic physical scales of these nanoscale chromatin PDs, enabling the evaluation of cancer-promoting markers in the nucleus for cancer screening or treatment care. Nucleus segmentation is a crucial task to ensure accurate analysis of chromatin PDs and spatial analysis, such as assessment of peripheral regions and nucleus shape. Traditional manual nuclei segmentation is labor-intensive, time-consuming, and prone to user bias, leading to errors due to over or under-segmentation. We present a novel hybrid convolutional neural networks-transformer-based approach for automatic cell nuclei segmentation with enhanced accuracy and speed for csPWS microscopy. Our method learns both local and global features from the input images for precise cell nuclei segmentation. We trained our model with several hundred csPWS images (>15,000 nuclei) from HCT116 and U2OS cell lines. With weighted-focal and dice loss, our model achieved median IoU and DSC scores of 0.77 and 0.89 for HCT116 cells and 0.75 and 0.85 for U2OS cells. This automated, rapid, and accurate segmentation streamlines csPWS microscopy data analysis, significantly aiding chromatin research for cancer diagnosis and treatment. Our deep-learning technique can also be leveraged by other microscopy techniques requiring segmentation, broadening its potential across various optical imaging modalities in biomedical research.