Title: Astra: A Universal Foundation Model for Generating 3D Computed Tomography Reports

Abstract:

The process of interpreting Computed Tomography (CT) scans is labor-intensive, requiring radiologists to scrutinize hundreds of volumetric slices for each case. This reliance on extensive manual review makes report generation both time-consuming and heavily dependent on specialized expertise. While automated systems for CT report generation hold significant potential for enhancing clinical efficiency, the domain currently lacks a foundational model capable of generalizing effectively. Specifically, there is an absence of a robust solution that supports multi-region reporting and maintains reliability across diverse, real-world external cohorts. This limitation stems from inherent variations in reporting styles and diagnostic terminology between different datasets, which causes naive joint training to suffer from noisy textual supervision, thereby restricting the model's generalizability.

To address these challenges, we introduce Astra, a generalizable foundation model designed for CT report generation. Astra was trained on a comprehensive dataset comprising 90,678 thoracoabdominal CT-report pairs (CTRgDB), covering 353,671 abnormalities across eight distinct organ systems. By standardizing report styles and employing reinforcement learning to enhance diagnostic consistency, Astra delivers reports that are both stylistically uniform and diagnostically precise across various anatomical areas and institutions.

When evaluated against CTRgDB and six external cohorts, Astra demonstrated state-of-the-art performance, achieving a 44.1% average improvement in fine-grained diagnostic metrics (P<0.001). In practical clinical settings, the integration of Astra into workflows resulted in a 29.6% acceleration in the drafting of chest reports and an 11.3% increase in the completeness of abdominal reports (P<0.001). Beyond immediate clinical assistance, Astra exhibits broad applicability as a foundational tool for CT AI development. It enhances downstream diagnostic capabilities and facilitates the scaling of vision-language pretraining through the synthesis of high-quality reports. Ultimately, Astra functions as an accessible clinical aid and serves as critical infrastructure for the future of AI-driven healthcare.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC