DALE-CT: Depth-Aware Foundation Models for Computed Tomography

A research team has introduced DALE-CT, a family of 2D vision models for computed tomography that achieves near-parity with state-of-the-art 3D vision-language systems while using significantly less data and no textual supervision, according to a paper submitted in 2026 [1]. The model family, called Depth-Aware Latent-Euclidean Computed Tomography, was built entirely from scratch using the Latent-Euclidean Joint-Embedding Predictive Architecture, or LeJEPA [1]. The researchers from the Institute for Biomedical Informatics Center for Applied AI at the University of Kentucky trained the models on the CT-RATE dataset and compared performance against a continually pre-trained DINOv2 baseline [1][4]. To improve representation quality, the team developed a 3D depth-aware pre-training strategy that uses dense auxiliary supervision from automated anatomical masks and human-annotated abnormalities [1]. Under linear probe evaluation with Multiple Instance Learning for multi-abnormality detection, the frozen backbone of the dual-supervised variant, DALE-CT-2S, achieved a Macro AUROC of 0.833 [1]. The authors state this result reaches near-parity with leading 3D vision-language models [1]. The DALE-CT family includes several variants. The base model, DALE-CT-0, was trained purely with self-supervised LeJEPA objectives and no auxiliary supervision, making it suitable as a general-purpose feature extractor for tasks such as segmentation and anomaly detection [4]. It uses a ViT-Large architecture with a patch size of 16 and native 512x512 resolution, supporting variable input sizes dynamically [4]. The DALE-CT-1S variant adds a single auxiliary objective, predicting the presence of 118 anatomical classes from TotalSegmentator within cropped regions [5]. Both models were initialized from scratch with no pre-trained weights [4][5]. The training objective for the base model combines a spatial invariance loss with Sketched Isotropic Gaussian Regularization, a technique that projects embeddings onto random 1D directions to enforce normality without traditional predictor networks [4]. The work explores 2D slice-based architectures as a flexible alternative to native 3D models for processing volumetric CT data [1]. By using dynamic depth-aware slab sampling and targeted auxiliary supervision, the framework avoids the need to aggressively resize or interpolate variable clinical scans [3]. The network instead learns continuous anatomical trajectories directly from sequential 2D slices [3]. All training code, evaluation scripts, and model weights have been made publicly available to support reproducibility [1][4][5]. The models are released under a CC BY-NC-SA 4.0 license, inherited from the CT-RATE dataset terms [4].