Glo-In-One

The glomerulosclerosis subtypes a nephropathologist most wants counted are the ones a classifier almost never sees. Across 22,077 annotated glomerular patches from 157 kidney biopsy patients, 6,647 glomeruli were obsolescent — but only 735 were solidified and 459 disappearing, and annotating more slides leaves the rare classes rare. Glo-In-One's answer is to find glomeruli where labeling is free: the figures of the published pathology literature, harvested at scale for self-supervised pretraining rather than supervision.

The toolkit wraps that idea in a pipeline returning three things a pathologist can act on: circle detections of every glomerulus, as XML that loads into Aperio ImageScope; a binary mask per glomerulus; and a lesion class from a closed vocabulary of normal, obsolescent, solidified, disappearing, and non-glomerular. A non-programmer runs the whole three-step path — detection, patch classification, segmentation — from the command line over a .svs file.

It was built by Yuankai Huo's group at Vanderbilt University with renal pathologists at Vanderbilt University Medical Center and a collaborator at the Big Data Institute of Central South University, and published in the Journal of Medical Imaging in 2022. The group later extended it with Glo-In-One-v2, which keeps the detection stage and replaces the whole-tuft mask with fourteen intraglomerular tissue and lesion classes; GloPath pushes the same entity-centric pretraining further, on over a million glomeruli from in-house biopsies rather than web figures.

#Key Features

• Circle representation for detection: The detection stage is the group's anchor-free CircleNet, predicting a center heatmap, a local offset, and a radius — fewer degrees of freedom to fit than a box, which suits ball-shaped objects like glomeruli.
• Classification as a false-positive filter: The classifier's explicit non-glomerular class is fed back to prune spurious detections, so characterization improves detection rather than merely consuming it.
• Self-supervised pretraining from web images: Over 10,000 compound figures retrieved from NIH Open-i are split into subplots by a compound-figure separation network, yielding more than 30,000 unannotated glomeruli across light, electron, and fluorescence microscopy.
• Detect-then-segment: Each detected glomerulus is re-cropped at full scanning resolution with 50 pixels of padding before masking, so segmentation never sees the downsampled slide.
• Released unlabeled corpus: The web-mined collection is published on Mendeley Data under CC BY 4.0 as a reusable pretraining set.

#Technical Details

Pretraining uses SimSiam over a ResNet-50 backbone on the 30,000 web-mined glomeruli, images resized to 224×224 and trained with momentum SGD at a base learning rate of 0.05 and batch size 64 under cosine decay. For characterization the backbone is frozen and one linear layer is trained with focal loss. On the in-house five-class cohort under five-fold patient-disjoint cross-validation, it reaches 68.0 balanced accuracy and 62.5 F1 with all labels, against 65.0 and 61.7 for a ResNet-50 trained from scratch on that same full label set; with 10% of the labels it still reaches 65.1. Applied without finetuning to an external public cohort of 2,340 glomeruli (1,170 normal, 1,170 sclerosed), it scores 94.5 AUC against 92.4 for an ImageNet-initialized model and 91.3 for training from scratch.

Detection reaches 0.627 average precision with the classifier acting as a filter, up from 0.621 for CircleNet alone, and 0.951 AP50. Segmentation uses DeepLab v3 for a patch-wise Dice coefficient of 0.955 ± 0.045 at 512×512, ahead of a U-Net at the same resolution, trained on 704 patches from 42 slides and tested on 147 internal plus 385 external ones. Slides are PAS-stained and scanned at 40× with 0.25 µm per pixel, and training ran on a single NVIDIA Quadro P5000.

#Applications

The intended user is a renal pathologist or nephrology researcher with a stack of biopsy slides and no programming support. Because the output is per-glomerulus, it supports the counting work behind biopsy reporting and cohort studies — sclerosed fractions, lesion-subtype burden, tuft area — and the XML overlay drops into the viewer pathologists already use. The authors point to kidney transplant assessment, where glomerulosclerosis characterization matters and manual counting does not scale.

#Impact

Glo-In-One demonstrates that unlabeled images scraped from the published literature transfer to a clinical histology task, and that the gain is largest exactly where annotation is scarcest: it beats supervised training on the solidified and disappearing classes while using a fraction of the labels. Its limits are stated plainly by the authors — the overall improvement over supervised baselines is around one point, per-class accuracy on the rare sclerosis subtypes sits near 50%, and no public multi-class cohort existed to test generalization beyond the binary case. The five-class head is closed and no embedding API is exposed, making this a fixed-function pipeline rather than a backbone to build on. Code is Apache-2.0, but the three checkpoints ship through a personal cloud-storage folder with no checksums or versioned release, and only the web-mined portion of the training data is public.