Glo-In-One-v2

The structures a nephropathologist cares about inside a glomerulus are nested, not disjoint: the tuft sits within Bowman's capsule, the mesangium sits within the tuft, and a globally sclerotic lesion blankets almost the whole capsule. A segmentation network that gives every class its own output channel treats those as fourteen unrelated problems and throws away the containment relationships — one reason lesion segmentation, where targets are small and annotator agreement is poor, has lagged behind whole-glomerulus segmentation.

Glo-In-One-v2 answers that with a single network carrying one dynamic segmentation head whose convolution kernels are generated per class at inference time, so all fourteen classes share a backbone and the contextual information that flows through it. It was built by Yuankai Huo's group at Vanderbilt University with renal pathologists at Vanderbilt University Medical Center and MD Anderson Cancer Center, posted as a preprint in November 2024 and published in the Journal of Medical Imaging in 2025.

It succeeds the group's earlier Glo-In-One toolkit, which detected glomeruli and segmented whole tufts but resolved nothing inside them; the detection stage carries over and the boundary-only segmentation is replaced. The same partially-labeled dynamic-head strategy appears in CLIP-Driven Universal Model for abdominal CT, where frozen text embeddings rather than a one-hot task code supply the class conditioning.

#Key Features

• Fourteen-class glomerular readout: Five intraglomerular tissue and cell classes — Bowman's capsule, glomerular tuft, mesangium, mesangial cells, podocytes — plus nine lesion types: global, segmental and nodular sclerosis, hyalinosis, capsular drop, adhesion, mesangial expansion, mesangial lysis, and microaneurysm.
• One dynamic head instead of fourteen models: A class-aware controller concatenates a one-hot task code with globally pooled image features and emits a three-layer segmentation head's parameters, so one network learns shared context across classes that overlap or contain one another.
• Training on partially labeled slides: Each annotated patch carries labels only for the class it was drawn for, so the task-conditioned formulation pools heterogeneous annotation efforts into one training set rather than one dataset per class.
• Hybrid cross-species transfer: Rather than training on mouse and applying to human zero-shot, rodent lesion annotations are mixed into training as an auxiliary source, lifting accuracy on human lesion classes too sparsely annotated to learn from alone.
• One-command containerized pipeline: A Docker image takes a directory of raw whole-slide images, runs detection then segmentation, and writes multi-channel masks — no training step, no Python required.

#Technical Details

The backbone is a residual U-Net; a class-aware controller fuses global average pooled features with the one-hot class vector to produce the three convolution kernels of the dynamic head, applied to the decoder output to yield that class's mask. Training used 23,529 annotated glomerular patches from 368 renal whole-slide images — 16,943 annotated directly by renal pathologists and 6,586 drawn from the Kidney Precision Medicine Project spatial segmentation dataset — cropped at full scanning resolution, resized to 512×512, and split 6:1:3 at the patient level. Batches of four were drawn from an image pool sized to the class count, over 200 epochs on a single NVIDIA RTX A5000.

Averaged over all fourteen classes the model reaches 76.5% Dice, against 58.4% for per-class DeepLabV3 models and 56.8% for per-class U-Nets and for Swin UNETR. Performance is stratified: tissue classes score near 90% or above (Bowman's capsule 96.3%, tuft 97.0%, mesangium 89.5%, global sclerosis 93.6%) while harder lesions trail (mesangial lysis 57.1%, adhesion 59.6%). On human lesion segmentation, hybrid rodent-and-human training reaches 70.4% average Dice, above 63.4% for zero-shot rodent-to-human transfer and 67.9% for human-only training; adding the tissue labels back in, despite their superset relationship to the lesions, raises it to 71.2%.

#Applications

The intended user is a renal pathologist or nephrology researcher with a stack of biopsy or preclinical slides and no appetite for manual outlining. Running end to end from whole-slide images, it supports quantitative phenotyping at cohort scale: podocyte and mesangial cell counts, mesangial area fractions, and per-lesion burden across every glomerulus on a slide. The cross-species coverage matters for translational work, where mouse models of glomerular disease are read with the same measurements later applied to human biopsies.

#Impact

Glo-In-One-v2 demonstrates that pooling annotations across species and incompletely labeled cohorts beats training a separate specialist per structure, and that fine-grained intraglomerular analysis can be packaged for non-programmers. Its limits are real: the label set is closed at fourteen classes with no route to a pathology it was not trained on, no general-purpose embedding is exposed, and lesion accuracy remains modest enough that outputs need pathologist review rather than standing alone. The code is GPL-2.0 on GitHub, but the trained weights ship through a personal cloud-storage link with no versioned release, and the Vanderbilt portion of the training data requires an institutional data use agreement, leaving only the Kidney Precision Medicine Project component publicly reproducible.