Computational and manual evaluation of SpinX shows high accuracy for spindle and cell cortex segmentation. (a) Violin plots show the distribution of IoU scores calculated from predictions with U-Net, Mask R-CNN, Cellpose, and SpinX-optimized. White marker within the box refers to the median, the shaded area refers to the estimated kernel probability density, and the box indicates the interquartile range of the data. Gray and red dots correspond to IoU scores smaller or >0.5, respectively. (b) Representative images show a range of different IoU scores calculated between the ground truth (red line) and predicted mask (yellow line) for the spindle (left) and cell cortex (right). Scale bars: 10 µm. (c) Representative SpinX prediction images for spindle (left) and cell cortex (right) describing the manual error classification system. Incorrectly segmented images were classified into “under segmentation minor” (U-Minor), “under segmentation major” (U-Major), “over segmentation minor” (O-Minor), “over segmentation major” (O-Major), and “multiple objects with artifacts” (MO). Insets show higher magnification of observed errors (yellow box). The prediction is highlighted by the blue and red overlays with the corresponding ground truth marked by a red dashed outline. Scale bars: 10 µm, 5 µm for inset. (d) Bar chart shows SpinX’s final accuracy, manually evaluated, for the spindle (white) and cell cortex (black) models. (e) Bar chart shows the proportion of incorrectly segmented images for each error type defined in c without Stage 3 of SpinX. For b, d, and e, N = 1,260 images (630 images each for spindle and cell cortex) from 10 3D time-lapse movies across four independent experiments were considered.