Skip to main content
Back to timeline
medRxivSource publication:

Same brown dye, two proteins: MorphoStain splits Tau from amyloid in Alzheimer's tissue, hitting 40.56 dB on 9,908 test patches and under 12 seconds per whole slide

Synopsis

In Alzheimer's pathology, Tau and amyloid are often marked with the same brown chromogen, so their spectra nearly coincide in RGB images and stain separation becomes mathematically ill-posed. MorphoStain embeds Beer–Lambert optical-density constraints in a latent diffusion teacher, then uses morphological routing to split the overlapping brown signal into amyloid, Tau and hematoxylin channels, and distills the teacher into a 0.12M-parameter student. On 9,908 held-out patches from the SEA-AD cohort it reaches 40.56 dB PSNR and 0.9961 SSIM, above the 38.84 dB of the unsupervised baseline InverseAE.

Source-provided article image: MorphoStain: Physics-Guided Latent Diffusion and Morphological Distillation for Multiplex Stain Decomposition in Alzheimer's Neuropathology
Fig. 1 · medRxiv · Page 2

Interpretation

In dual-stain multiplex immunohistochemistry, amyloid and Tau share the brown DAB chromogen and are nearly spectrally indistinguishable; MorphoStain separates that overlap into independent amyloid, Tau and hematoxylin channels. The prior unsupervised reference, InverseAE, lacks morphological routing and suffers severe crosstalk bleed when resolving overlapping brown DAB into Tau and amyloid channels. On 9,908 held-out patches from the SEA-AD cohort, MorphoStain reaches a mean reconstruction PSNR of 40.56 dB and SSIM of 0.9961 versus 38.84 dB for InverseAE, a 1.72 dB gap with paired-test p<10⁻¹⁶ and Cohen's d=3.52.

Separation needs no pixel-level manual annotation: a physics-constrained diffusion teacher generates pseudo-labels that are distilled into a small network able to run in real time on whole slides. Diffusion models can capture multimodal posteriors and report uncertainty, but iterative per-patch sampling is too slow across gigapixel whole-slide images, and unconstrained sampling can violate image-formation physics. The teacher is conditioned on Phikon pathology foundation-model features and runs five reverse steps with five perturbed realizations each on the Beer–Lambert optical-density manifold, producing mean concentrations and epistemic variance; the FastStudentU-Net takes a 5-channel input of RGB concatenated with the physical concentration projection and shrinks from 169.67M to 0.12M parameters, over 1,400× compression.

The distilled student keeps reconstruction quality while pushing inference to real-time, so a full gigapixel whole slide can be unmixed and stitched in just over ten seconds. Iterative diffusion sampling on the teacher side cannot serve clinical throughput, and the student replaces stochastic sampling with convolutional smoothing. The student reports 41.66 dB PSNR and 174.6 FPS (5.73 ms latency) on the calibration subset, while the main benchmark lists MorphoStain at 140.1 FPS and states it can process a whole slide larger than 50,000×40,000 pixels in under 12 seconds.

Perspective

The demonstrated scope is 512×512 tiles at 20× from 10 middle temporal gyrus whole slides of 5 SEA-AD donors, with training, validation and test split strictly by donor; zero-shot tests use 100 brain patches from external donor H14.09.010 and breast-cancer HER2/hematoxylin sections. The current setting targets a two-chromogen (M=2) DAB and hematoxylin system, and the authors propose next embedding differentiable absorbance estimation into Phikon features for slide-specific stain adaptation and wiring uncertainty maps into scanners for real-time human-in-the-loop triage.

The authors name three operational limits: extreme over-staining may exceed nominal optical-density bounds; the pipeline currently relies on upstream tile rejection, with uncertainty maps not yet wired directly into scanners; and the method covers only M=2, with extension to M≥4 multiplex immunofluorescence left as a direction. The uncertainty head marks out-of-focus and tissue-folding regions with u_var>0.0025 where PSNR falls below 35 dB, a threshold drawn from inspection of the observed anomalous regions.