Loading…
Virtual staining network that turns label-free multiphoton brain-tissue images into H&E and Perls Prussian Blue histology, from unpaired data.
No providers recorded yet. Browse all providers
A neuropathologist handed a multiphoton image of brain tissue is reading a picture nobody trained them on. Second-harmonic generation and two-photon excited fluorescence render collagen, elastic fibers and hemosiderin as pseudo-colored intensity channels rather than the pink and blue of hematoxylin and eosin, and in a blinded comparison three neuropathologists confused hyaline-degenerated vessels with collagen-aged ones on multiphoton images far more often than on stained slides. That legibility gap keeps label-free microscopy out of the workflows where its advantages — no fixation, no dye, no wait on a stain — would count for most.
CCM-stainGAN closes the gap by translating the multiphoton image into the stains pathologists already read. Developed at Fuzhou University, Fujian Normal University, the Fujian Medical University hospitals and Southern Medical University, it maps two-channel multiphoton images to virtual H&E and three-channel images to virtual Perls Prussian Blue, the iron stain that reveals hemosiderin. It is the virtual-staining component of qMPM, a quantitative multiphoton workflow for cerebral cavernous malformation — a lesion whose repeated microhemorrhage causes drug-resistant epilepsy and whose surgical outcome turns on how completely the perilesional hemosiderin rim is removed.
Multiphoton and stained images can never be pixel-registered: the stain goes on an adjacent 10 μm section, so the same cells are never in the same places. CCM-stainGAN therefore starts from unpaired cycle-consistent translation, then adds constraints on tissue identity — an unconstrained generator can satisfy the cycle while quietly rendering a hyaline-degenerated vessel as a normal one. The same lab later took the reverse direction with ViFIT, generating fiber modalities from H&E.
The generator is a U-Net-style encoder-decoder joined by skip connections, with residual blocks at each scale to keep gradients flowing and a classification head reading the bottleneck features; the released training configuration uses nine tissue-component classes. Discrimination uses a PatchGAN, which scores local patches and so rewards high-frequency stain texture. Training runs end-to-end in PyTorch with Adam at an initial learning rate of 0.0002 and exponential decay of 0.999, for 100 epochs with decay from epoch 50, on a single NVIDIA RTX 3090.
The data come from 35 frozen sections cut at 10 μm from 12 patients with CCM-related epilepsy, imaged on a Zeiss LSM 880 with a tunable Ti:sapphire laser at 810 nm excitation and 30 mW average power. Emission is split into an SHG channel at 395–415 nm, a TPEF-1 channel at 428–570 nm for NADH and FAD, and a TPEF-2 channel at 600–695 nm for hemosiderin-related products; adjacent sections stained with H&E, Elastic van Gieson and Perls Prussian Blue serve as reference. In the blinded reading study — 10 specimens, 333 validation images, three neuropathologists reading multiphoton, real H&E and virtual-stained images a week apart — the virtual stains reached 98.8% accuracy on CCM histopathological features, the highest of the three, and recovered the hyaline-versus-aging distinction that multiphoton images alone lose. Against a CycleGAN baseline, the added constraints preserve more histopathological structure, most visibly in stitched large-scale images.
The target setting is intraoperative and postoperative neuropathology for cavernoma-related epilepsy, where the surgeon must decide how far beyond the visible lesion to cut. Because multiphoton imaging needs no staining protocol, a fresh section can be imaged and virtually stained in less time than digital H&E scanning requires, and the Prussian Blue output presents hemosiderin — the feature most associated with seizure recurrence — in its familiar form.
The durable contribution is the classification constraint: an explicit tissue-component agreement term on unpaired stain translation, rather than trusting cycle consistency alone to preserve diagnosis-bearing structure. The code is MIT-licensed and ships with sample multiphoton, H&E and Prussian Blue images, though no trained checkpoint is distributed, so reproducing the model means running the training script. Evidence is single-center and small — 12 patients, one lesion type, one microscope — and the model is fitted to CCM tissue rather than histopathology in general, so performance on other specimens or other multiphoton systems has not been established.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.