Universal CT Representations from Anatomy to Disease Phenotype through Agglomerative Pretraining

2026-05-22 · Source: cs.CV updates on arXiv.org · Field: Science & Research — Health & Medical Research, Artificial Intelligence & Machine Learning, Mathematics & Computational Sciences · Depth: Expert, extended

Summary

FlexiCT is a family of CT foundation models developed through agglomerative continual pretraining on 266,227 CT volumes from 56 publicly available datasets. This large-scale public resource supports slice-level, volume-level, and vision-language analysis across three stages: 2D axial, 3D anatomical, and report-guided semantic alignment. FlexiCT matches or exceeds prior task-specific approaches on 18 benchmarks across five downstream task families, including segmentation, classification, registration, vision-language understanding, and clinical retrieval. Its embeddings further organize CT scans along gradients associated with various tumor stages, suggesting that CT foundation models can capture imaging features relevant to disease phenotype characterization. The code is available at https://github.com/ricklisz/FlexiCT.

Key takeaway

For AI Scientists and Machine Learning Engineers developing medical imaging solutions, FlexiCT offers a unified foundation model approach that significantly reduces the need for task-specific models. You should consider integrating FlexiCT to accelerate deployment, especially for rare conditions or new protocols where annotated data is scarce, leveraging its label efficiency and emergent cross-modal capabilities.

Key insights

Agglomerative pretraining unifies diverse CT analysis tasks and captures disease severity.

Principles

• Sequential pretraining preserves complementary features.
• Self-supervised features encode genuine anatomical structure.
• Agglomerative strategy enhances label efficiency.

Method

FlexiCT uses three-stage agglomerative continual pretraining: 2D axial, 3D anatomical, and report-guided semantic alignment, leveraging DINOv3, iBOT, and contrastive losses.

In practice

• Use FlexiCT for label-efficient disease classification.
• Apply FlexiCT-2D for training-free cross-modal registration.
• Leverage FlexiCT-3D-VLM for zero-shot disease identification.

Topics

• CT Foundation Models
• Agglomerative Pretraining
• Medical Image Segmentation
• Cross-Modal Registration
• Vision-Language Models
• Disease Phenotype Characterization
• Label-Efficient Learning

Code references

• ricklisz/FlexiCT

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Emerging Flexible Designs for Geospatial Multimodal Foundation Models — Apples-to-apples comparison of geospatial foundation model architectures reveals design trade-offs in flexibility, alignment, and performance.
• On-Manifold Variational Learning with Heat-Kernel Priors — Manifold-anchored variational learning improves medical image representation by keeping prototypes on the data manifold.
• Disease-Centric Vision-Language Pretraining with Hybrid Visual Encoding for 3D Computed Tomography — The framework improves 3D CT diagnosis by integrating hybrid visual encoding, disease-level contrastive learning, and diagnosis-aware prompting.
• BreastGPT: A Multimodal Large Language Model for the Full Spectrum of Breast Cancer Clinical Routine — Workflow-aligned data and cross-scale visual modeling are critical for clinically grounded medical MLLMs.
• Scalable Training of Spatially Grounded 2D Vision-Language Models for Radiology — Spatially grounded radiology VLMs can be trained at scale without manual annotations, maintaining language quality.
• PRET is a few-shot system for pan-cancer recognition without example training — PRET enables flexible, scalable, and effective pan-cancer recognition with minimal examples, outperforming traditional AI and human pathologists.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.