MICAD 2026 · 7th International Conference on Medical Imaging and Computer-Aided Diagnosis

DINO-Nucleus

Multi-Level Self-Supervised Features for Nucleus Instance Segmentation

Mohammad Sarhangzadeh · Elyar Esmaeilzadeh · Doruk Öner

NeuraVision Lab · Department of Computer Engineering · Bilkent University

Read the paper Contact GitHub code · Coming soon

Abstract

Nucleus instance segmentation and counting are essential for quantitative microscopy, enabling reliable analysis of cell populations, tissue organization, and disease-related changes. However, accurately separating touching nuclei remains difficult, and many existing approaches rely on computationally expensive models or sensitive post-processing procedures. We present DINO-Nucleus, an efficient framework that leverages multi-level self-supervised DINOv2 representations for nucleus instance segmentation and counting. The model jointly predicts nucleus regions, pixel-wise instance flows, and explicit nucleus centers, while a consistency objective encourages the predicted flows to converge toward their corresponding centers. Multi-level transformer features provide complementary spatial and semantic information, and the pretrained backbone is adapted using lightweight LoRA modules to reduce fine-tuning cost. Experimental results show that DINO-Nucleus achieves strong segmentation and counting performance while remaining computationally efficient. These findings demonstrate that parameter-efficient adaptation of multi-level self-supervised features can provide an effective solution for nucleus instance segmentation in microscopy images.

01 / OVERVIEW

Full architecture

DINO-Nucleus overall architecture, three prediction heads, training objective, and instance reconstruction — **Figure 1.** DINO-Nucleus architecture, training objective, and instance reconstruction.

02 / METHOD

Method and training objectives

Learning nucleus regions, centers, and pixel-wise directions from fused DINOv2 features.

DINO-Nucleus combines intermediate representations from DINOv2 ViT-S/14 (layers 2, 5, 8, and 11) with high-resolution CNN features. A shared decoder makes three predictions, each supervised by a corresponding objective.

Foreground loss

L foreground equals L binary cross-entropy plus L Dice.

Binary cross-entropy and Dice loss learn the nucleus foreground region.

Center loss

L center equals Gaussian focal loss of predicted center heatmap and target center heatmap.

Gaussian focal loss supervises a center heatmap, built from interior nuclei centers.

Flow regression loss

Foreground-masked Smooth L1 supervises the predicted directions within nucleus regions.

TARGET

Center-directed flow target

Each foreground pixel points to an interior nucleus center; background flow is zero.

TRAJECTORY MECHANISM · SCHEMATIC

Following the flow toward nuclei centers

04 / ENDPOINT CONSISTENCY

Flow-to-center consistency

Local flow regression alone does not guarantee that trajectories reach valid centers. We differentiably integrate each foreground pixel's predicted flow, then encourage its endpoint to lie at a nucleus-center response.

Differentiable trajectory integration

At each time step, the position is updated by normalized predicted flow weighted by foreground support and step size.

Center reference with a detached prediction

The reference center map combines the ground-truth center map with a detached predicted center map.

Endpoint-consistency loss

The endpoint loss is the foreground-weighted negative log of the reference center response at trajectory endpoints.

Combined objective

Total loss is a weighted sum of foreground, center, flow, and endpoint-consistency losses.

Loss weights: 1.0 foreground · 0.5 center · 0.5 flow · 0.05 endpoint, with a warm-up schedule for consistency.

03 / FEATURES

Inside DINOv2 representations

Figure 2: PCA feature maps, nucleus prototype scores, and t-SNE plots for DINOv2 layers 2, 5, 8, and 11 — **Figure 2.** PCA maps, nucleus–background prototype scores, and t-SNE projections of background, boundary, and nucleus-dominant patches.

04 / EXPERIMENTS

Results on BBBC038

670 microscopy images · instance matches at IoU > 0.45.

Method	Dice ↑	IoU ↑	Obj. F1 ↑	PQ ↑	AJI ↑	MAE ↓	Pearson ↑
TC-UNet++	92.46	88.51	83.64	72.33	77.59	9.19	89.70
Cellpose3	88.71	81.34	88.52	76.69	83.13	5.97	93.39
StarDist	85.94	81.07	83.06	70.05	75.93	9.72	87.62
CellViT	90.22	84.35	90.75	77.40	85.35	5.63	94.58
DINOv2	90.98	86.16	83.51	74.88	79.28	7.47	89.20
DINO-Nucleus	90.45	85.30	92.38	79.33	88.97	3.11	98.75

Numbers are reported in Table 1 of the MICAD 2026 paper; they have not yet been reproduced using the new codebase. Scores follow the paper's scale.

05 / CONCLUSION

Conclusion

DINO-Nucleus brings together transferable self-supervised representations and explicit geometric instance modeling. By combining multi-level DINOv2 features with local image detail and lightweight LoRA adaptation, the framework learns nucleus structure without requiring full backbone fine-tuning.

Joint foreground, center, and flow predictions—linked by an endpoint-consistency objective—allow neighboring nuclei to be separated through their learned trajectories rather than watershed reconstruction. The experiments suggest that this design is particularly effective for instance separation and reliable nucleus counting, offering a practical and parameter-efficient approach to quantitative microscopy.

CITATION

BibTeX

@inproceedings{sarhangzadeh2026dinonucleus,
  title     = {DINO-Nucleus: Multi-Level Self-Supervised Features
               for Nucleus Instance Segmentation},
  author    = {Sarhangzadeh, Mohammad and Esmaeilzadeh, Elyar
               and Oner, Doruk},
  booktitle = {7th International Conference on Medical Imaging
               and Computer-Aided Diagnosis (MICAD)},
  year      = {2026}
}

Provisional conference citation. Publication details such as pages and DOI can be added once available.