Research map / Detection and segmentation / CNN features

Beyond ViT Tokens: Masked-Diffusion Pretrained Convolutional Pathology Foundation Model for Cell-Level Dense Prediction

NeurIPS 2026 · Detection and segmentation › CNN features · approach: Convolution-transformer backbones · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CMD, a convolutional ConvNeXt-UNet pretrained with pixel-space masked diffusion conditioned on frozen H0-mini/UNI features, used as a frozen dense feature extractor for cell-level pathology segmentation.

Problem

Cell-level dense prediction in pathology (nuclear segmentation) where ViT-based pathology foundation models lose fine local morphology due to patch tokenization, and dense annotations are costly.

Approach

A fully convolutional ConvNeXt-UNet pretrained by masked-diffusion (timestep-controlled structural masking with L1 reconstruction) in pixel space, with frozen pathology foundation model features injected via adaLN, used as a frozen dense feature extractor with lightweight heads.

Key results

• Frozen CMD-L with a 21.21M-parameter head reaches Dice 0.889 on CPM-15, 0.880 on CPM-17 and 0.854 on TNBC, above the frozen ViT foundation models listed (e.g. UNI 0.738, Hibou 0.794 on CPM-15).
• On few-shot TNBC 1-shot, frozen CMD-L Dice is 0.714 versus 0.497 for frozen UNI2-h and 0.498 for ViT-Adapter.
• In the backbone comparison, ConvNeXt-UNet gives BF1 33.66 on CPM-15 versus 19.65 for DiT under linear probing.

Surprising findings

• Pixel-space masked diffusion outperforms VAE latent-space and high-resolution VAE variants on the dense benchmarks.
• Frozen CMD-L with a 7M-parameter head beats end-to-end ViT-Adapter (364M trainable) in the 1-shot CPM-17 and TNBC settings.

Limitations

• Only three to five small nuclear segmentation datasets evaluated; no broader pathology tasks reported despite the abstract's 'multiple tasks'.
• Relies on frozen UNI/H0-mini conditioning, so gains partly depend on an existing ViT-based PFM.
• Scaling analysis is limited to two model sizes and pretraining steps, not a full scaling law.

Experimental setup

Datasets: HistAI unlabeled pathology collection CPM-15 CPM-17 Janowczyk Kumar TNBC

Benchmarks: CPM-15 CPM-17 TNBC Janowczyk Kumar

Metrics: Dice Precision BF1 95% bootstrap confidence intervals

Baselines: PathOrchestra CONCH CONCHv1.5 Prov-GigaPath Hibou H-Optimus-0 H-Optimus-1 Kaiko Midnight-12k MUSK Phikon Phikon-v2

Builds on: masked diffusion DDPM ConvNeXt TransUNet ViT-Adapter

Closest papers by meaning

• Dynamic Entity-Masked Graph Diffusion Model for Histopathology Image Representation Learning (AAAI 2025 · similarity 0.88)
• Semantic and Visual Crop-Guided Diffusion Models for Heterogeneous Tissue Synthesis in Histopathology (NeurIPS 2025 · similarity 0.88)
• Towards Scalable Context-Aware Single-Cell Spatial Transcriptomics Prediction from Histology Images (NeurIPS 2026 · similarity 0.87)
• Why Invariance is Not Enough for Biomedical Domain Generalization and How to Fix It (NeurIPS 2026 · similarity 0.86)
• FOCUS: Knowledge-enhanced Adaptive Visual Compression for Few-shot Whole Slide Image Classification (CVPR 2025 · similarity 0.86)
• PathVQ: Reforming Computational Pathology Foundation Model for Whole Slide Image Analysis via Vector Quantization (NeurIPS 2025 · similarity 0.86)