MRPT

A gigapixel whole-slide image (WSI) carries diagnostic signal at several scales at once: nuclear morphology at 40x, glandular and stromal organization at 20x, and tissue architecture at 10x. Most computational pathology foundation models are pretrained on patches cropped at a single magnification, which forces them to learn one scale well and generalize to the others by accident. Patch-level encoders such as UNI and CONCH lack whole-slide spatial context altogether, while slide-level models such as Prov-GigaPath and TITAN recover global context but flatten the nested structure of the slide.

Multi-Resolution Pyramid Transformer (MRPT) addresses both gaps at once. Developed at Khalifa University of Science and Technology with University of Western Australia and posted as a preprint in August 2026, it builds a slide representation by climbing a hierarchy — cells to patches to regions to the full slide — while simultaneously fusing three magnifications at every rung. It extends the same group's MR-PLIP, which introduced multi-resolution pretraining at the patch level but did not exploit the slide's hierarchical organization.

The mechanism that ties the two axes together is Consecutive Cross-Resolution Attention (CCRA). Rather than letting every magnification attend to every other, MRPT restricts cross-attention to adjacent pairs — 10x with 20x, 20x with 40x — mirroring how pathologists zoom through intermediate magnifications instead of jumping from overview to cell.

#Key Features

• Consecutive cross-resolution attention: Class tokens from two adjacent resolutions are averaged and appended to the finer sequence, then used as a single query against it. Because the query is one vector, the operation is linear rather than quadratic in sequence length.
• Four-level hierarchy: Co-registered regions decompose into patches and then into cell-level tokens, and each level's encoder is frozen and reused as the embedding layer for the level above it.
• Frozen general-purpose extractor: The pretrained encoders are evaluated across 34 public datasets through zero-shot prompting, linear probing, and attention-based multiple instance learning, with no per-dataset retraining.
• Compact backbone: With small ViT backbones at each stage the full model holds fewer than 12M parameters, and downstream tuning touches as little as 0.5M.
• Conversational extension: MRPT-LLaVA couples the slide representation to an LLM through cross-modal alignment and visual instruction tuning for slide-level question answering and report generation.

#Technical Details

Each slide is read at 10x, 20x, and 40x. Regions of 4096x4096 pixels at 20x, with their spatially aligned 2048x2048 and 8192x8192 counterparts, are split into 256 patches, which are split again into 256 cell-level tokens per resolution. Pretraining runs in three self-supervised stages with teacher-student networks: a cell-level CCRA transformer over 624M patches, a patch-level encoder over 2.4M regions, and a slide-level encoder over 36K WSIs — 30,000 from TCGA and 6,000 from CPTAC. The three resolution branches use 2, 4, and 4 encoder layers respectively. Training and inference ran on four NVIDIA A100 GPUs.

On whole-slide linear probing MRPT averages 0.898 balanced accuracy against 0.845 for TITAN; on zero-shot slide classification across 10 datasets it reaches 0.806 against 0.751. Patch-level linear probes average 92.75% versus 86.12% for the next-best encoder, and attention-based MIL over MRPT features averages 0.863 balanced accuracy against 0.812 for MR-PLIP and 0.798 for UNI. MRPT-LLaVA averages 80.77% accuracy across four slide-level VQA benchmarks. Linear-probe inference costs 5.2 minutes per slide, comparable to the 4.3 and 4.4 minutes reported for Prov-GigaPath and TITAN.

#Applications

The intended use is as a frozen feature extractor for slide-level pathology tasks that have too few labeled slides to train an encoder from scratch. Evaluations span cancer subtyping on TCGA-NSCLC, TCGA-RCC, TCGA-BRCA, PANDA, UBC-OCEAN, and BRACS, metastasis detection on Camelyon16 and Camelyon17, tissue phenotyping, and — through MRPT-LLaVA — visual question answering and automated report generation. Because the encoder emits embeddings at cell, patch, region, and slide granularity, downstream users can attach a linear classifier, an MIL aggregator, or a language model without retraining the backbone.

#Impact

MRPT's contribution is architectural: it shows that hierarchy and magnification are separable axes, and that constraining attention to adjacent scales is both cheaper and more accurate than fusing all scales at once. That the resulting model is smaller than most pathology foundation models while outperforming them on the reported benchmarks makes the design attractive for groups without large GPU budgets. The work is a preprint awaiting peer review, its comparisons are the authors' own, and while the paper announces a code and model release, no repository or weights have been published, so independent reproduction is not yet possible.