Research Square · Sep 9, 2026

Chest2Vec: A multipurpose text encoder conditioned by instructions for chest radiograph and computed tomography reports

By Ko H, Yang R, Lee D et al.

Text encoders turn radiology reports into finding labels, retrieval targets, and image-text supervision, making them core infrastructure for chest imaging artificial intelligence. However, existing encoders target general or biomedical language, or chest radiographs, and none has been developed for chest computed tomography (CT) reports. We propose Chest2Vec, a multipurpose, instruction-conditioned text encoder for chest radiograph and chest CT reports at 0.6 and 4 billion (B) parameters. One frozen encoder per size served retrieval, error discrimination, finding classification, and anatomy-specific representation with only the instruction changed and a linear head where required; the 0.6B encoder supplied image-text supervision. On chest CT reports from a held-out hospital, Chest2Vec-4B retrieved the paired impression in 68.8% versus 57.1% for the strongest comparator, LLM2Vec4CXR, and reached 0.692 versus 0.478 macro F1 by linear probing. A radiograph-specialized encoder remained stronger on external radiograph retrieval. We release the encoders, structured chest CT data, and report labeler.

doi.org/10.21203/rs.3.rs-10799556/v1Open accessAI & data

AgingResearch.News