Histopathology image translation model that standardizes H&E staining style, then generates virtual collagen, reticulin, and trichrome fiber images.
No providers recorded yet. Browse all providers
The fibrous scaffold around a tumour carries diagnostic information that hematoxylin and eosin barely shows. Collagen alignment, reticular fiber loss, and elastic fiber disruption mark invasion and stromal remodelling, but reading them requires either a special stain consumed on a separate section or label-free second-harmonic generation microscopy on equipment few pathology departments own. ViFIT, developed across Fuzhou University, Fujian Normal University, and two Fujian Medical University teaching hospitals, generates those fiber views computationally from the H&E slide already on the microscope.
The difficulty is that virtual staining models are brittle to the input they receive. H&E appearance varies with reagent lot, protocol, and scanner, not only between hospitals but between batches inside one lab, and a fiber generator trained on one appearance misreads another — most often by confusing nuclei with fibers. Rather than absorb that variance into a single end-to-end network, ViFIT splits the problem in two: an unsupervised standardization stage maps any incoming H&E appearance to a pathologist-preferred reference style, and a second stage translates the standardized image into fiber modalities. Constraining the input distribution sharply reduces how much paired fiber data the second stage needs.
Both stages are unsupervised, which matters because paired H&E and fiber images cannot be pixel-registered: the fiber stain comes from an adjacent section, so the cells never line up. Published in Medical Image Analysis after a January 2025 preprint, it approaches collagen from a different direction than iQMAI, which regresses collagen structure from H&E against polarization imaging of Picrosirius Red slides.
Stage one is a cyclic generative adversarial network with two identically structured but separately parameterised ResNet generators — a Standardization Generator mapping toward the reference style and a Reconstruction Generator mapping back — trained with CNN discriminators, a cycle consistency loss, and the two pretext tasks. The style query becomes an extra input channel, and each generator emits three image channels plus a style-label channel. Stage two adds the auxiliary intensity-reversal generator, paused once the fiber generator produces usable coarse maps.
Training used a multi-style breast H&E dataset of 8,578 images spanning seven staining styles, drawn from Fujian Medical University Union Hospital with the public BreakHis, BACH, and BreCaHAD collections. The H&E-to-fiber dataset comprises 112 images, including whole slides, covering seven disease types from breast, lung and thyroid tumours to meningioma, with second-harmonic images acquired on unstained adjacent sections at 810 nm excitation and 30 mW average power. Standardization was scored with normalized mutual information, kernel inception distance, and multi-scale SSIM, where the closest competitor, HistAuGAN, trails by roughly 10%. On fiber transformation, ViFIT exceeds DRIT by 18% on virtual second-harmonic output and VirtualMultiplexer by 20% on virtual reticulin. Across 230 regions of interest analysed with CurveAlign, mean fiber density differs from the reference sections by 0.03, against deviations above 0.1 for comparison models.
The intended setting is routine surgical pathology. In postoperative cases the model surfaces tumour-associated fiber patterns and supplies alignment and orientation metrics that reduce subjective grading variability; in frozen-section intraoperative work, where special stains are impractical on the clock, it offers fiber context immediately. Standardization alone addresses a common referral problem — slides arriving at a tertiary centre in an unfamiliar staining style, which often forces resectioning and restaining. Because the fiber output feeds quantification tools directly, it also serves stromal remodelling research without a multiphoton system.
ViFIT is a narrow generative model rather than a general histopathology backbone, and its contribution is the two-stage decomposition: standardize appearance first, and virtual staining becomes tractable from a fiber dataset of about a hundred images. The repository ships downloadable weights and a scripted pipeline, so the model runs on new H&E images without retraining, though it carries no license file. Each fiber modality currently needs its own trained model, and generalization beyond the seven tissue types evaluated has not been established.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.