TITAN

TITAN (Transformer-based pathology Image and Text Alignment Network) is a slide-level multimodal foundation model for computational pathology developed by the Mahmood Lab at Harvard Medical School and Brigham and Women's Hospital, with collaborators at Helmholtz Munich and the University of Tokyo. Introduced in a November 2024 preprint and published in Nature Medicine in 2025, TITAN produces a single embedding for an entire gigapixel whole-slide image (WSI), addressing the patient- and slide-level questions that clinical pathology actually asks.

Most pathology foundation models — including the Mahmood Lab's own UNI and CONCH — operate at the level of small patches or regions of interest and require an external multiple-instance-learning aggregator to reach a slide-level decision. TITAN instead learns the aggregation directly: it consumes a grid of CONCHv1.5 patch features spanning a full slide and encodes them into a compact, transferable slide representation. Because the slide encoder is aligned to text during pretraining, the same embedding supports classification, retrieval, and generation without task-specific fine-tuning or clinical labels.

Critically, TITAN is pretrained without any manual annotations, learning entirely from unlabeled slides, matched pathology reports, and machine-generated captions. This makes it well suited to resource-limited settings such as rare-disease retrieval and prognosis, where labeled training data is scarce.

#Key Features

• Slide-level embeddings: TITAN encodes an entire WSI into one general-purpose vector, removing the need for a separate multiple-instance-learning head to move from patches to a slide-level prediction.
• Built on CONCHv1.5 patches: The slide encoder ingests 768-dimensional features extracted by the CONCHv1.5 patch encoder from non-overlapping 512x512 pixel patches at 20x magnification, reusing a strong pathology patch backbone.
• 2D ALiBi long-context encoding: A Vision Transformer with Attention with Linear Biases extended to two dimensions lets TITAN extrapolate to slide extents far larger than those seen during pretraining.
• Three-stage vision-language pretraining: Visual self-supervision is followed by contrastive alignment to region captions and then to full clinical reports, yielding an embedding space shared between histology and text.
• Label-free multitask capability: Without fine-tuning, TITAN supports linear probing, few-shot and zero-shot classification, rare-cancer and cross-modal retrieval, and pathology report generation.

#Technical Details

TITAN's slide encoder is a Vision Transformer with approximately 48.5 million parameters using 2D ALiBi positional encoding. Pretraining proceeds in three stages. Stage one applies iBOT self-supervised learning to 16x16 feature grids corresponding to 8,192x8,192 pixel regions (roughly 4x4 mm at 20x) drawn from 335,645 WSIs. Stages two and three use the CoCa contrastive-captioning objective to align slide and region representations first to 423,122 synthetic captions generated by the PathChat pathology assistant, then to 182,862 slide-report pairs from clinical text. The training corpus, Mass-340K, combines de-identified Mass General Brigham slides with public GTEx data. Across linear-probing, few-shot, zero-shot, retrieval, and report-generation benchmarks, TITAN reported performance exceeding both region-of-interest and prior slide-level foundation models.

#Applications

TITAN is designed as a slide-level feature extractor and multimodal engine for computational pathology research. Its embeddings drive cancer subtyping, biomarker and molecular-status prediction, and slide-level prognostic modeling with minimal labeled data. The vision-language alignment enables cross-modal search — retrieving morphologically similar cases from a text query or surfacing relevant report text for a slide — which is especially valuable for rare cancers with few labeled examples. TITAN can also draft pathology report text directly from slide features, supporting research into automated reporting.

#Impact

TITAN advances computational pathology from patch- and region-level encoders toward general-purpose slide-level foundation models that fold aggregation and language grounding into a single pretrained system, complementing the Mahmood Lab's UNI and CONCH. Code and gated weights are released on GitHub and HuggingFace, lowering the barrier to slide-level research. Access is restricted: the weights carry a CC BY-NC-ND 4.0 research-only license and are gated behind registration, the multimodal decoder was withheld to reduce the risk of leaking protected health information, and the Mass-340K training corpus is private and reflects the patient population of a single US academic medical center. TITAN has not been validated for clinical diagnostic use and requires independent evaluation before any diagnostic application.