Highlights:
- PLUTO-4 introduces frontier-scale foundation models for digital pathology.
- Features two architectures: PLUTO-4S for efficiency and PLUTO-4G for maximal precision.
- Trained on over 551,000 whole-slide images (WSIs) from more than 50 institutions.
- Achieves an 11% performance boost in dermatopathology diagnosis tasks.
TLDR:
PLUTO-4 represents a breakthrough in computational pathology, combining large-scale training and transformer-based architectures to deliver state-of-the-art diagnostic performance across diverse histopathology tasks. Its scalable models set a new benchmark for AI-driven medical imaging and diagnostic research.
A research team led by Harshith Padigela (https://arxiv.org/search/cs?searchtype=author&query=Padigela,+H), along with Shima Nofallah (https://arxiv.org/search/cs?searchtype=author&query=Nofallah,+S), Atchuth Naveen Chilaparasetti (https://arxiv.org/search/cs?searchtype=author&query=Chilaparasetti,+A+N), Ryun Han (https://arxiv.org/search/cs?searchtype=author&query=Han,+R), Andrew Walker (https://arxiv.org/search/cs?searchtype=author&query=Walker,+A), Judy Shen (https://arxiv.org/search/cs?searchtype=author&query=Shen,+J), Chintan Shah (https://arxiv.org/search/cs?searchtype=author&query=Shah,+C), Blake Martin (https://arxiv.org/search/cs?searchtype=author&query=Martin,+B), Aashish Sood (https://arxiv.org/search/cs?searchtype=author&query=Sood,+A), Elliot Miller (https://arxiv.org/search/cs?searchtype=author&query=Miller,+E), Ben Glass (https://arxiv.org/search/cs?searchtype=author&query=Glass,+B), Andy Beck (https://arxiv.org/search/cs?searchtype=author&query=Beck,+A), Harsha Pokkalla (https://arxiv.org/search/cs?searchtype=author&query=Pokkalla,+H), and Syed Ashar Javed (https://arxiv.org/search/cs?searchtype=author&query=Javed,+S+A), has unveiled PLUTO-4: Frontier Pathology Foundation Models. This innovative system extends the Pathology-Universal Transformer (PLUTO) to new frontiers of scale and capability, offering cutting-edge tools for computational pathology, a rapidly evolving field that merges artificial intelligence with medical imaging.
PLUTO-4 introduces two transformative Vision Transformer (ViT) architectures: PLUTO-4S and PLUTO-4G. The compact PLUTO-4S variant is designed for efficient, multi-scale deployment using a FlexiViT setup enhanced with 2D-RoPE (Rotary Positional Embedding). It delivers high throughput, low-latency processing for clinical or research environments where speed and adaptability are critical. In contrast, the frontier-scale PLUTO-4G harnesses a single high-resolution patch size to achieve exceptional representational depth and training stability. Both models are pre-trained using a self-supervised objective inspired by DINOv2, enabling powerful feature extraction from vast pathology datasets without the need for manual annotation.
A cornerstone of PLUTO-4’s success is its extensive training corpus—over 551,000 whole-slide images (WSIs) representing 137,144 patients and more than 60 disease types, gathered from across 50 institutions. This massive, heterogeneous dataset ensures robust generalization across staining protocols, tissue types, and clinical settings. The evaluation reports demonstrate that PLUTO-4 achieves state-of-the-art results in multiple key pathology tasks, including patch-level classification, segmentation, and full-slide diagnosis. Notably, the PLUTO-4G model exhibits an 11% performance improvement in dermatopathology diagnosis benchmarks, underscoring its significance for real-world diagnostic applications. With its scalable architecture and high transferability, PLUTO-4 holds promise as a universal backbone for translational pathology research, predictive modeling, and AI-assisted diagnostics in healthcare systems worldwide.
Source:
Source:
arXiv:2511.02826 [cs.CV] — ‘PLUTO-4: Frontier Pathology Foundation Models’ by Harshith Padigela, Shima Nofallah, Atchuth Naveen Chilaparasetti, Ryun Han, Andrew Walker, Judy Shen, Chintan Shah, Blake Martin, Aashish Sood, Elliot Miller, Ben Glass, Andy Beck, Harsha Pokkalla, and Syed Ashar Javed. DOI: https://doi.org/10.48550/arXiv.2511.02826