CircleNet

A glomerulus is round, and so is a nucleus. A bounding box is not. Because a biopsy section can be mounted and scanned at any angle, the tightest box around a round object changes its width and height as the slide turns, though the object has not. CircleNet's answer is to change the target rather than the network: predict a center point and a single radius — a bounding circle — and the shape parameters drop from two degrees of freedom to one.

The framework is anchor-free and built on the CenterNet detection family. A convolutional backbone feeds three parallel heads: a heatmap head that localizes object centers, a local offset head that recovers sub-pixel position, and a circle radius head that regresses one radius per center under an L1 loss. There are no anchor boxes, no region proposals and no keypoint-grouping stage, so a slide patch is processed in a single forward pass. To score circle predictions against circle ground truth, the authors define circle IoU (cIOU) from the closed-form area of intersection between two discs; under random displacements of the test annotations it tracks conventional IoU almost exactly, licensing its use as the matching criterion in mean average precision.

CircleNet was developed by Yuankai Huo's computer science group at Vanderbilt University, with renal pathologists at Vanderbilt University Medical Center and a collaborator at PAII. It appeared at MICCAI 2020 on glomerulus detection and was extended in IEEE Transactions on Medical Imaging in 2021 with nuclei detection. The same lab reuses it as the detection stage of Glo-In-One.

#Key Features

• Bounding circle as the detection target: One radius describes a ball-shaped object that a box describes with width and height, and it is unchanged by rotating the image.
• Circle IoU: A closed-form overlap metric for two discs, so evaluating circle detections needs no conversion back to boxes.
• Rotation consistency as a reported metric: The fraction of detections that still overlap after the test image is rotated 90 degrees is measured directly.
• Whole-slide inference into a pathology viewer: A packaged script reads a .scn scan and writes an XML annotation file that overlays the detections in Aperio ImageScope.
• Released checkpoints with a small inference API: A two-line call returns a list of [x, y, radius, confidence, category], with no training step on the user's own data.

#Technical Details

Hourglass-104 and DLA-34 serve as backbones, both initialized from COCO-pretrained CenterNet weights and fine-tuned for 50 epochs on a single 11 GB GPU. The glomerulus cohort comes from renal biopsies scanned at 0.25 µm per pixel and downsampled to 4 µm per pixel: 42 biopsies contributing 704 glomeruli for training, 7 with 98 for validation and 7 with 147 for testing, expanded to 7,040 training, 980 validation and 1,470 testing patches of 512×512 pixels. With the DLA backbone CircleNet reaches 0.647 average precision against 0.598 for CenterNet-DLA and 0.584 for Faster R-CNN with ResNet-50, at AP50 of 0.907 and AP75 of 0.787, and a rotation consistency of 0.886 against 0.851.

Nuclei detection uses the public MoNuSeg 2018 challenge data — 30 H&E images of 1000×1000 pixels containing 21,623 annotated nuclear boundaries drawn from TCGA slides. Here the Hourglass backbone leads at 0.487 mAP against 0.447 for CenterNet-HG, with rotation consistency of 0.891 against 0.793. The circle loses ground on medium-sized objects, roughly 2% of that test set, where nuclei are elongated rather than round; the authors identify an ellipse with a second degree of freedom as the natural remedy.

#Applications

The immediate use is quantitative renal pathology: locating and counting glomeruli across a biopsy so that sclerosed fractions and cohort-scale phenotyping do not depend on manual annotation. Because the output is a center and a radius rather than a mask, it also serves as the first stage of detect-then-segment pipelines, where each detected object is re-cropped at full scanning resolution before a segmentation network sees it — the arrangement Glo-In-One uses. The nuclei checkpoints extend the same idea to H&E computational pathology, where nuclear counts and density underpin tumor grading.

#Impact

CircleNet established circle representation as a practical alternative to the bounding box wherever biomedical objects are approximately round: the idea has been lifted to a sphere for lung nodule detection in 3D CT, and the same lab later built circle-contour instance segmentation on top of it. Its scope is deliberately narrow — a closed class set, one detection task, no embedding interface, no demonstrated transfer to classes it was not fine-tuned on. The code is MIT-licensed and borrows from CenterNet, but the glomerulus training set is not public, the checkpoints are distributed as Vanderbilt Box links without checksums or a tagged release, and the maintainers note that the original model files were lost and the published weights are reproductions whose results may vary slightly.