M^3-GloDet: Benchmark for Diseased Glomeruli

• The paper introduces a systematic benchmarking framework for multi-region and multi-scale instance segmentation of fine-grained diseased glomeruli in renal pathology.
• It evaluates various detection architectures by comparing metrics like average Dice scores and mAP across different patch sizes and magnification levels.
• The benchmark provides practical insights for optimizing patch extraction and scale selection, guiding future development of adaptive, scale-aware AI models in digital pathology.

Searching arXiv for the primary paper and closely related renal pathology / instance segmentation references. M$^3$-GloDet, written in the paper title as M$^3$-GloDets, is a systematic benchmarking framework for multi-region, multi-scale instance segmentation and detection of fine-grained diseased glomeruli in renal pathology. It is designed to evaluate how detection and segmentation architectures behave when glomeruli are imaged at different magnifications and extracted with different patch sizes, addressing a setting that had been comparatively understudied relative to normal glomeruli or global sclerosis. Rather than introducing a single new detector, the framework establishes a rigorous evaluation protocol spanning multiple diseased subtypes, multiple architectures, and controlled variations in region size and image scale (Shi et al., 25 Aug 2025).

1. Task definition and clinical scope

M$^3$-GloDet studies the detection and segmentation of diseased glomeruli in whole-slide kidney pathology images. The target task is instance segmentation: each glomerulus must be localized with a detection output such as a bounding box and delineated with a pixel-level mask. This is explicitly more difficult than semantic segmentation because individual glomeruli must be separated even when they are small, clustered, or morphologically similar (Shi et al., 25 Aug 2025).

The framework focuses on five clinically relevant categories: GGS (global glomerulosclerosis), Viable Glom (viable glomerulus), Ischemic Glom (ischemic glomerulus), SGS (segmental glomerulosclerosis), and Atubular Glom (atubular glomerulus). The paper characterizes these categories as subtle and heterogeneous, which makes them difficult for standard detectors. It also situates the benchmark against an imbalance in prior work: research efforts had largely emphasized normal glomeruli or global sclerosis, while fine-grained diseased subtype detection and instance-level delineation were comparatively underexplored (Shi et al., 25 Aug 2025).

This framing is important because the benchmark is not merely about object presence. It is about whether computational pathology systems can resolve subtype-specific morphology under realistic whole-slide imaging conditions. A plausible implication is that performance in this setting is constrained not only by detector design but also by how much contextual tissue is visible and at what optical resolution that context is represented.

2. Multi-region and multi-scale design

The central premise of M$^3$-GloDet is that glomerular morphology is sensitive to both region-of-interest size and imaging magnification. Whole-slide images are too large to process directly, so the benchmark formalizes patch extraction as a controlled experimental variable. At fixed 20× magnification, the multi-region experiment evaluates patch sizes of 2048 × 2048, 4096 × 4096, and 8192 × 8192 pixels (Shi et al., 25 Aug 2025).

This patch-size axis encodes a trade-off. Small patches provide less context but potentially sharper lesion focus and lower compute. Large patches provide more anatomical context but also more background noise, higher memory demand, and possible dilution of lesion detail. M$^3$-GloDet tests these effects systematically rather than treating patching as a preprocessing detail (Shi et al., 25 Aug 2025).

The second axis is magnification. At a fixed patch size of 4096 × 4096 pixels, the multi-scale experiment compares 20×, 10×, and 5×. The motivation is that high magnification captures fine structure, whereas lower magnification may suppress distracting details and improve generalization. The paper explicitly tests the hypothesis that higher detail is not always better (Shi et al., 25 Aug 2025).

Within this formulation, the “M$^3$” denotes the paper’s emphasis on multi-region and multi-scale analysis. The benchmark therefore operationalizes a practical methodological question for digital renal pathology: scale selection is treated as part of model evaluation rather than an external nuisance parameter.

3. Dataset construction and annotation protocol

The benchmark dataset comprises 8 nephrectomy cases collected retrospectively at Northwell Health—specifically Long Island Jewish Medical Center and Northshore University Hospital—from July 2017 to October 2019 (Shi et al., 25 Aug 2025). The slides contain non-cancerous kidney parenchyma, are PAS-stained, were digitized using Leica GT450 RUO scanners, and were scanned at 20× magnification.

Glomeruli were first identified in QuPath, after which expert renal pathologists manually annotated them. Each glomerulus received both a pixel-level instance mask and a class label from the five disease categories. The split is performed at the slide level: Train: 5 slides, Validation: 1 slide, and Test: 2 slides (Shi et al., 25 Aug 2025).

Split	GGS	Viable Glom	Ischemic Glom	SGS	Atubular Glom
Train	1338	520	211	233	79
Validation	154	18	51	17	15
Test	140	231	54	66	10

These counts show strong class imbalance, with Atubular Glom as the rarest class. The paper also notes that class imbalance coincides with morphology heterogeneity, which becomes visible in later category-wise performance analysis. This suggests that the benchmark is not only a scale-selection study but also a stress test for rare and structurally inconsistent lesion types.

4. Benchmarked architectures and evaluation protocol

M$^3$-GloDet compares seven representative segmentation and detection architectures under consistent computational settings and using their official architectures on an NVIDIA A6000 Ada GPU (Shi et al., 25 Aug 2025). The evaluated models span both long-standing baselines and newer state-of-the-art systems:

• Mask R-CNN: ResNet-50-FPN backbone; two-stage detector with RPN and RoI alignment.
• Cascade R-CNN: ResNet-50-FPN backbone; multi-stage detector with progressively higher IoU thresholds.
• SOLOv2: ResNet-50-FPN with deformable convolutions; single-stage, grid-based instance segmentation.
• DiffusionInst: Swin-Base backbone; diffusion-based instance segmentation using vectorized mask representation.
• ASF-YOLO: CSPDarknet53 backbone; attention plus scale fusion for cell instance segmentation.
• CelloType: Swin-Large backbone; end-to-end joint segmentation and classification with DINO and MaskDINO.
• YOLOv12: R-ELAN backbone; attention-centric real-time detector.

The benchmark uses three metrics. The Dice coefficient measures mask overlap and is defined as

$\mathrm{Dice} = \frac{2|P \cap G|}{|P| + |G|},$

where $P$ is the predicted mask and $G$ is the ground truth. Macro-F1 is computed across the five categories using an IoU threshold of 0.5. mAP is computed over IoU thresholds

$^3$0

following COCO-style evaluation (Shi et al., 25 Aug 2025).

The paper further notes that Dice was reported differently depending on analysis: in tables, as a single score per category on the full test set; in boxplots, as per-patch Dice scores excluding empty patches. This distinction matters because per-patch distributions expose heteroskedasticity across classes that aggregate averages can obscure.

5. Main empirical findings

In the multi-region experiment, the key conclusion is that 4096 × 4096 patches offered the best compromise between context and efficiency. According to the paper, this size provides enough surrounding tissue context without introducing too much irrelevant background, losing structural cues, or becoming computationally unwieldy. It was the most consistent “sweet spot” across models (Shi et al., 25 Aug 2025).

Among the evaluated methods, YOLOv12 achieved the strongest overall segmentation performance. At 4096-pixel patches, it obtained Average Dice = 52.23% and Average F1 = 50.65%. It also achieved the best Dice for Atubular Glom, with 39.08%, which is notable because this class is both rare and difficult (Shi et al., 25 Aug 2025). By contrast, CelloType consistently achieved the highest mAP, indicating particularly strong localization quality even when mask overlap quality was comparatively weaker. The paper characterizes Cascade R-CNN and SOLOv2 as relatively balanced middle-tier methods, Mask R-CNN as stable but not top-performing, and DiffusionInst and ASF-YOLO as underperforming in this dataset. Some configurations of DiffusionInst and CelloType could not be run at the largest patch size because of memory constraints.

In the multi-scale experiment, the central result is that maximum magnification is not always optimal. The paper reports that 20× provides the most detail, 10× often gives the best balance of detail and context, and 5× can sometimes generalize better by smoothing away noisy fine detail (Shi et al., 25 Aug 2025). For Mask R-CNN and Cascade R-CNN, 10× often offered a good trade-off. More strikingly, YOLOv12 achieved its best overall performance at 5× in this experiment, with Average Dice = 52.45%, Average F1 = 52.07%, and Average mAP = 35.62%. The paper explicitly interprets this as evidence that resolution should be matched to architecture rather than maximized blindly.

At the category level, GGS and Viable Glom were generally easier, with higher medians and more stable Dice distributions. Ischemic Glom and SGS were harder, with lower medians and greater variability. Atubular Glom was the hardest overall due to rarity and morphological diversity. The boxplot analysis shows relatively compact performance distributions for common classes and wide variability for rare and structurally inconsistent classes, indicating that class imbalance and morphology heterogeneity are major obstacles (Shi et al., 25 Aug 2025).

6. Interpretation, practical implications, and terminological clarification

The principal contribution of M$^3$1-GloDet is methodological. The paper presents it as the first systematic benchmarking framework dedicated to diseased glomerular detection with instance-level evaluation for multiple diseased subtypes, cross-architecture comparison across classical and modern detectors, explicit study of region size and magnification, and practical guidance for digital renal pathology workflows (Shi et al., 25 Aug 2025). Its findings therefore concern both model ranking and experimental design.

Several practical implications follow directly from the reported results. Automated renal pathology systems should not be developed or evaluated at a single fixed magnification or patch size. Intermediate region sizes should be considered for balancing context and compute. Moderate magnifications may improve generalization. Strong class imbalance, especially for rare lesion types, should be expected. Architecture selection should depend on the clinical objective: the paper associates CelloType with stronger localization through higher mAP, and YOLOv12 with the best overall balance of segmentation, detection, and efficiency (Shi et al., 25 Aug 2025).

The paper also points toward future work in adaptive scale-aware architectures, improved patching strategies, models that jointly optimize robust localization and accurate mask prediction, and broader studies with more cases and more diverse disease spectra (Shi et al., 25 Aug 2025). This suggests that M$^3$2-GloDet is best understood as a benchmark and experimental framework for scale-aware renal pathology AI rather than as a monolithic algorithmic proposal.

A potential source of confusion is the reuse of the label “M$^3$3” in unrelated literatures. In particular, “M$^3$4: Dense Matching Meets Multi-View Foundation Models for Monocular Gaussian Splatting SLAM” addresses monocular SLAM, dense correspondence estimation, and Gaussian splatting scene reconstruction, not renal pathology or glomerular detection (Ren et al., 17 Mar 2026). The shared notation should therefore not be taken to imply methodological continuity. In the renal pathology context, M$^3$5-GloDet refers specifically to multi-region and multi-scale benchmarking for diseased glomerular instance segmentation (Shi et al., 25 Aug 2025).