CONCHv1.5

A whole-slide image tiled at the conventional 256x256 pixels yields tens of thousands of patches, and every slide-level model downstream inherits that sequence length. CONCHv1.5 is the Mahmood Lab's answer to that arithmetic: a histopathology patch encoder trained to operate on 512x512 pixel tiles at 20x magnification, which quarters the number of tokens a slide produces. The bet is that a larger tile carries proportionally more tissue per vector, so the shorter sequence costs nothing in representation quality; the authors credit the robustness of patch-level pathology encoders at higher input resolutions.

It is an extended second generation of CONCH, and its lineage crosses both of the lab's encoder lines. The vision tower is a ViT-L/16 restored from the UNI self-supervised pathology checkpoint rather than trained from scratch, then trained with the CoCa contrastive-captioning objective that CONCH introduced. What most users touch is only the image half: titan.return_conch() returns the frozen vision tower and its evaluation transform, and that is the interface the CLAM and TRIDENT preprocessing toolkits call.

CONCHv1.5 was released on 2 December 2024 alongside TITAN, the slide-level model built on its features, and it is documented in TITAN's Nature Medicine study rather than in a paper of its own. It has since become the default patch backbone across the lab's stack — Threads and the CONCH v1.5 variant of ABMIL both consume its features frozen.

#Key Features

• Higher-resolution tiling: Encoding 512x512 pixel tiles at 20x instead of the widely used 256x256 reduces a slide's feature-grid sequence length by a factor of four, which is what makes long-context slide encoders tractable.
• Initialized from a pathology backbone: Starting the ViT-L trunk from an existing self-supervised histology checkpoint rather than from random weights lets the vision-language stage refine an already pathology-aware representation.
• Preparation-robust pretraining: The corpus spans formalin-fixed paraffin-embedded sections, frozen tissue and immunohistochemistry, so features hold up across laboratories with different processing protocols.
• Compact 768-dimensional output: A single-query attentional pooler and layer norm project the 1,024-wide trunk down to 768 dimensions per tile, keeping slide-scale feature grids small enough to store and reload.
• Frozen drop-in extractor: It is selectable by name in CLAM (conch_v1_5) and TRIDENT (conch_v15), so swapping patch backbones in an existing pipeline is a configuration change rather than a code change.

#Technical Details

The released artifact is the vision tower: a ViT-L/16 trunk with a 1,024-dimensional width, followed by an attentional pooler with eight heads and a single contrastive query that maps to a 768-dimensional embedding, then a layer norm. The distributed checkpoint is roughly 1.2 GB in FP32, about 307 million parameters. Tiles cut at 512x512 are resized to 448x448 at inference and the position embeddings are interpolated from the pretraining grid. Training used the CoCa objective over 1.26 million image-caption pairs for 20 epochs; the accompanying text encoder and multimodal decoder are each 12 transformer layers with an embedding dimension of 768 and a hidden dimension of 3,072, and TITAN reuses both pretrained rather than training its own. The published record evaluates CONCHv1.5 as TITAN's feature extractor rather than isolating it on a patch-level benchmark, but it appears throughout the study's baselines: simple mean pooling over CONCHv1.5 tile features is used as the reference unsupervised slide representation, and on the 46-class TCGA-OncoTree pan-cancer cohort its embeddings separate organ-level clusters that two dedicated slide encoders fail to resolve.

#Applications

CONCHv1.5 is a frozen feature extractor for whole-slide image pipelines. The standard workflow segments tissue, tiles it at 512x512 and 20x, and caches one 768-dimensional vector per tile; those vectors then feed multiple-instance-learning aggregators for subtyping, biomarker status and survival modeling, or a slide encoder for a single slide-level embedding. Because pretraining covered immunohistochemistry and frozen sections as well as routine H&E, it is a reasonable default for cohorts where staining and processing vary, including intraoperative frozen sections and IHC panels. The coarser tiling helps most when the downstream model must attend across a whole slide.

#Impact

CONCHv1.5 marks a shift in how pathology patch encoders are judged: not on their own region-of-interest leaderboards, but on how well they compress a gigapixel slide for whatever consumes it next. Its 512-pixel tiling is now a design assumption in several downstream models, and its integration into CLAM and TRIDENT makes it reachable from pipelines that have no other connection to TITAN. The caveats are real. Weights are gated behind HuggingFace registration requiring an institutional email and Mahmood Lab approval, and the CC-BY-NC-ND 4.0 license restricts use to non-commercial academic research and forbids redistribution. Published characterization is thin — a handful of sentences in another model's Methods, with hyperparameters deferred to an extended data table — and no patch-level ablation isolates what the second generation gained over the first. CONCHv1.5 has not been validated for clinical diagnostic use.