Fig 1.
Image with ground truth segmentation.
Nucleus signal (a), cytoplasm signal (b), and ground truth segmentation (c).
Fig 2.
Feature pyramid fusion of nucleus features.
Pre-trained nucleus features (violet) are fused with features of the feature pyramid in the cell detection and segmentation model (green) by either concatenation or addition.
Table 1.
Reduced ResNet-50 architecture.
Fig 3.
Including nucleus information for cell segmentation.
(a) without nucleus information, (b) with additional input for the nucleus channel, and (c) with fused nucleus features.
Fig 4.
For each box proposal, crops are resized to 28 × 28 pixels. Crop from the input image (a), full-image segmentation mask (b), cell mask (c) and weight matrix (d).
Fig 5.
Visualization of cell segmentation errors on a 256 × 256 patch of clustered cells in test data: Predicted masks (red) differ only slightly from ground truth masks (white).
Table 2.
Detection and segmentation results for the nuclei test dataset in terms of AP.
Table 3.
Detection results for the cells test dataset in terms of AP.
Table 4.
Segmentation results for the cells test dataset in terms of AP.
Table 5.
Detailed cell segmentation results for an IoU threshold of 0.75 on the cells test dataset.
Table 6.
Detection results for clustered cells test dataset in terms of AP.
Table 7.
Segmentation results for clustered cells test dataset in terms of AP.
Fig 6.
APs for cell segmentation on clustered cells.
Fig 7.
Visualization of segmentation of clustered cells.
Top: nucleus signal (a), cytoplasm signal (b), and ground truth segmentation (c). Bottom: Instance segmentations predicted by for model without nucleus information (d), with nucleus channel (e), and FPF ⊕ with weighted loss (f).
Fig 8.
Cell segmentation of clustered cells by Feature Pyramid Fusion (FPF) on a 512 × 512 patch.