A Hierarchical Deep Learning Framework for Nuclei 3D Reconstruction from Microscopic Stack-Images of 3D Cancer Cell Culture

In this article, we propose a hierarchical deep learning framework for the nuclei 3D reconstruction from a stack of microscopic images representing 3D cancer cell culture. The framework goes through three successive stages namely: at the slice level of the stack (i) the spheroid detection and (ii) their nuclei segmentation then at the stack level (iii) the nuclei 3D reconstruction. For this purpose, we prepared a dataset of bright-field microscopic images acquired from 3D cultures of HeLa cells and manually annotated by the experts for both tasks (spheroids detection and nuclei segmentation). Two CNN models namely, YOLOv5x and U-Net-VGG19 have been trained and validated on our dataset for the detection and the segmentation tasks, respectively. For the 3D reconstruction task, the delaunay triangulation technique has been adopted by exploiting point cloud clusters that represent the segmented nuclei in the stack. Our framework offers to the biologists an efficient assisting tool for quantifying the number of spheroids and analyzing the morphology of their nuclei. The conducted experiments on our generated dataset show the promising results obtained by our framework with notably an average precision of 0.892 and 0.76 on the spheroids detection and nuclei segmentation respectively. Moreover, our 3D reconstruction technique shows visually a consistant representation of nuclei in term of volumetery and shape.