Learned compression autoencoders for histopathology whole-slide images, tuned so reconstructions preserve the features downstream models rely on.
No providers recorded yet. Browse all providers
Digital pathology generates enormous files. A single whole-slide image scanned at 20x magnification can run to gigapixels, and a hospital archive of them quickly becomes a storage and bandwidth problem rather than a research asset. JPEG shrinks the files, but it was tuned for natural photographs and human viewers, and at aggressive settings it erases the nuclear texture and stromal detail that diagnostic models read. The question this work asks is whether a learned compressor can hit the same file sizes while keeping the features that matter.
Pathology Fine-Tuned AEs is the family of released checkpoints answering that question, from the Computer Vision Lab at Stony Brook University with collaborators at Argonne National Laboratory, published at MICCAI 2025. Rather than train a compressor from scratch, the authors repurposed three autoencoders originally built as the latent-space front ends of text-to-image latent diffusion models — the Stable Diffusion 1.5 VAE, the Stable Diffusion 3.5 Large VAE, and DC-AE — and fine-tuned their decoders specifically on histopathology.
The critical design decision is what "fidelity" means. PSNR and SSIM, the standard image-quality metrics, correlate poorly with whether a pathology model can still do its job on a reconstruction. So the fine-tuning objective is instead a perceptual loss computed in the embedding space of UNI, a pathology foundation model: the decoder is trained to produce reconstructions that a pathology encoder sees as the original, not reconstructions that a pixel metric likes.
Fine-tuning used 24 million 256x256 patches at 20x magnification, extracted from 2,500 TCGA whole-slide images spanning breast, colon, and prostate, on 8 NVIDIA A100 GPUs for 120,000 iterations at a learning rate of 5e-5. The released checkpoints hold 83.7M parameters (SD-1.5 VAE), 83.8M (SD-3.5 Large VAE), and 323.4M (DC-AE). Reconstruction quality was scored as cosine similarity between original and reconstructed embeddings under UNI, Phikon-v2, and Prov-GigaPath. Fine-tuning raised UNI similarity from 0.733 to 0.906 for DC-AE, 0.837 to 0.932 for the SD-1.5 VAE, and 0.959 to 0.978 for the SD-3.5 VAE.
Downstream validation substituted compressed reconstructions for original images with all task models held fixed. On BCSS segmentation, Dice fell from 71.57 on originals to 71.20 using DC-AE with K-means quantization, under a 1% drop, while JPEG-10 at comparable file size collapsed to 60.56. On TCGA-BRCA subtyping, accuracy went from 94.89% on originals to 94.05% with the fine-tuned SD-3.5 VAE and 93.30% with fine-tuned DC-AE. Additional evaluation covered NCT-CRC patch classification, CRAG gland segmentation, and multiple-instance-learning slide-level tasks.
The immediate use is archival and transmission: institutions holding large slide repositories can store compressed latents or reconstructions at a fraction of the raw footprint and still run segmentation, patch classification, and multiple-instance-learning pipelines on them. Because the encoders are unchanged from their latent diffusion originals, the same latents also feed generative histopathology work — synthesis, inpainting, and latent-space modeling — without a separate encoding pass. Collaborative and federated settings benefit most, where slide data must cross network links that raw WSIs saturate.
This is a narrower artifact than a general-purpose pathology foundation model: it compresses and reconstructs rather than producing task-agnostic representations. Its contribution is methodological as much as practical — the demonstration that pixel-level image-quality metrics are the wrong yardstick for medical image compression, and that optimizing against a domain foundation model's embedding space is a better one. The three checkpoints are public and ungated on Hugging Face; the paper does not state a license for them, and no training or fine-tuning code has been released, so reproducing the fine-tuning procedure requires reimplementation from the described recipe.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.