COMB

A whole-slide image is a gigapixel array, far larger than a GPU can hold, so a virtual staining network cuts it into tiles and generates each one on its own. Stitched back together, the tiles show their seams: a straight line of color shift running through tissue that has no boundary there, nuclei halved at a tile border. A pathologist has to decide at every seam whether a discontinuity is biology or an artifact of how the picture was computed, and a segmentation model simply treats the seam as an edge.

COMB — the Consistency Memory Bank — gives each tile access to its neighbors without ever holding them in the same computation. As the generator scans the slide column by column, it writes each tile's encoder and decoder feature maps into a memory bank and reads them back at two points. The margins of the current tile are padded with the real boundary features of the tiles beside it instead of with zeros or a reflection, which restores the receptive field that tiling truncated. And the channel-attention weights computed from the isolated tile are mixed with the average weights of its neighborhood, which damps the statistical drift that makes one tile come out pinker than the next.

The framework comes from Dou Hoon Kwark and colleagues at the University of Illinois Urbana-Champaign and the National Center for Supercomputing Applications, and was accepted to MICCAI 2026. It belongs to the label-free line of virtual staining, where the input is the tissue's own spectral signal — mid-infrared absorption or stimulated Raman scattering — so the section survives intact for molecular assays. Where CCM-stainGAN and its peers established that this translation is learnable at patch scale, COMB addresses what happens when the patches are reassembled.

#Key Features

• Retrieval-based local padding: Tile margins are filled with boundary features fetched from surrounding tiles, so convolutions operate on continuous tissue signal rather than synthetic padding.
• Neighbor-aware channel attention: The CBAM channel map of a tile is fused with the mean map of its neighbors within a context radius, pulling anomalous local statistics back toward the regional norm.
• Sliding-window memory schedule: Inference and look-ahead caching share one continuous scan, and only the columns the context radius needs are retained, cutting cached-tile memory from quadratic in the slide's tile grid to linear in its height.
• Batch-normalized throughout: Consistency is enforced without the instance-normalization variants used for stain-to-stain transfer, which destabilize sparse, heterogeneous spectral inputs.

#Technical Details

The generator is the VSGD-Net attention U-Net with its input head widened to the spectral channel count — 10 mid-infrared wavenumbers or 5 stimulated Raman bands — paired with a multi-scale PatchGAN discriminator under an LSGAN objective, and a perceptual loss on a ConvNeXt-Tiny backbone fine-tuned on histopathology in place of VGG-19. Training used Adam at a learning rate of 2e-4, batch size 8, and 200k iterations, with the memory bank set to a mixing weight of 0.3 and a context radius of 2. The paired corpus is frozen prostate surgical resections imaged in both modalities and co-registered to H&E at 0.5 µm/pixel: roughly 60,000 IR-HE and 16,500 SRS-HE 512×512 patches, evaluated by 3-fold cross-validation with patient-level separation. The seamlessness margin is large: on SRS, TexTile rises from 0.4058 (VSGD-Net) and 0.3743 (CC-WSI-Net) to 0.7320 against a ground-truth 0.8453, and Focused Sobel Gradient falls from 0.2500 to 0.1054, while SSIM improves more modestly to 0.4995. The sliding window holds a 30,720 × 28,160 slide in 4.38 GB rather than 10.64 GB, at 262.8 seconds on an RTX 2080 Ti.

#Applications

The target workflow is a pathology service that images unstained sections with an infrared or Raman scanner and reads them as H&E, leaving the tissue available for downstream molecular work. Because the output is a full slide rather than a gallery of patches, it feeds automated analysis directly: a UNet++ tumor segmenter trained on real H&E scores 80.97 Dice and 234.50 HD95 on COMB output, against 81.25 and 224.45 on the real stain, with the improvement concentrated in the boundary-sensitive metric that tiling artifacts corrupt most.

#Impact

Seamlessness is the practical obstacle between label-free virtual staining and a slide a pathologist will sign out, and COMB treats it as an architectural problem rather than a post-hoc blending step — the memory bank is a runtime cache of neighbor features, so a trained generator is applied unchanged to any slide. Its scope is correspondingly specific: the evaluation is a single-institution prostate cohort, the training data are not public, and the input head is sized to a modality, so the framework transfers across label-free contrasts while a given checkpoint does not. The code carries a top-level MIT license, though many of its Python files keep NVIDIA's non-commercial CC BY-NC-SA 4.0 header, and no pretrained weights accompany it: running the model means training it on paired spectral and H&E slides of one's own.