Fig. 3. Impact of image resolution on deep learning segmentation accuracy.

a–f Results of 2D U-Net models trained for semantic segmentation (inner area and boundary) of E. coli (a) and C. elegans (d) at varying simulated resolutions. b Intersection over Union (IoU) distributions (semantic and binary masks) across scaling factors for E. coli, ranging from ×4 downsampling to ×4 upsampling. IoU was calculated both for semantic segmentation (average IoU of edge and inner regions) and the binary masks resulting from considering the edge and inner regions. Purple line (log scale): theoretical throughput calculated as the number of objects observable by a unit of time (τ) given a certain resolution. c E. coli diameter distributions from U-Net segmentations versus ground truth. e Same as in (b), IoU for C. elegans segmentation at varying resolutions. g IoU on the test image volume of 3D U-Net mitochondria binary segmentation in transmission electron microscopy (TEM) at varying resolutions. h Results visualizations corresponding to z-colour-coded-projections at varying axial/lateral resolutions. i, j MicroSAM²² with a single point-wise prompting strategy was used to obtain the instance segmentation of E. coli (i) and C. elegans (j) example images. We pointed to the same object in each rescaled image and ran the automatic instance segmentation pipeline. All the images matched model input dimensions (512×512 pixels) to avoid internal tiling or image re-scaling. The images that had smaller dimensions were padded with zeros. Square boxes correspond to the same physical area on each image. Scale bars correspond to a 100, 25, and 10 pixels; d 100 and 25 pixels; h 200 pixels; i, j 50 µm. Circles in b and e correspond to data points outside the quartiles of the IoU distribution.