Video dataset and DL architecture for spindle and cell cortex image segmentation. (a) Representative images show complex variations in illumination and marker intensities intrinsic to time-lapse movies of subcellular structures (cell cortex, brightfield [label-free]; chromosomes, H2B-GFP; mitotic spindle, mCherry-Tubulin, or SiR-Tubulin dye). Ideal images are shown within the blue box. Red and yellow arrowheads indicate the object of interest and interfering variation(s), respectively. Scale bar: 5 μm. (b) DL model architecture of SpinX. The model expands the pre-existing Mask R-CNN architecture (ResNet101, FPN, RPN, ROI Align, and FCN [He et al., 2018]), by introducing a third stage (blue box). In Stage 1, the network performs object detection followed by the segmentation of spindle and cell cortex in Stage 2. Stage 3 (highlighted in blue) links temporal and spatial information in 3D live-cell movies through tracking and generates a consistent mask of the same object through time. The inputs of the model are grayscale or RGB images of various sizes (5D input). The outputs are binary masks of the same size as inputs with predicted foreground regions, bounding box coordinates (rectangular boxes in teal and red, Stage 1) and the corresponding Class ID (Stage 2). Scale bar: 10 µm.