ReadCT: Parameter-Efficient 3D Chest CT Findings Generation Using Low- Rank Adaptation
Related Experiment Video
Updated: Nov 6, 2026
Development and Evaluation of 3D-Printed Cardiovascular Phantoms for Interventional Planning and Training
Published on: January 18, 2021
Maram Mahmoud A Monshi1, Fahad Mahmoud Monshi2
1Department of Information Technology, College of Computers and Information Technology, Taif University, Taif, Saudi Arabia.
Introduction/Objective:
Automatic report generation from Three-Dimensional (3D) chest Computed Tomography (CT) is computationally demanding. We evaluated whether a compact, parameter-efficient model could generate findings text directly from 3D non-contrast chest CT volumes within a tractable retrospective benchmark setting.
Methods:
ReadCT encodes each standardized 3D volume using a lightweight 3D convolutional network, converts the global representation into compact visual-prefix tokens, refines the visual-text sequence with a single-layer Transformer encoder, and generates findings with a single-layer Long Short-Term Memory (LSTM) decoder. Low-Rank Adaptation (LoRA) was applied to selected projection layers while the embeddings and the decoder remained trainable. We trained and evaluated ReadCT on large-scale public chest CT resources, using CT-RATE (more than 25,000 scans with reports) as the primary benchmark. Performance was assessed using standard lexical and semantic report-generation metrics, and generalization was examined on an external dataset (BIMCV-R) without additional tuning.
Results:
On the CT-RATE held-out test split, ReadCT achieved a Bilingual Evaluation Understudy (BLEU)-4 score of 29.17 and a BERTScore F1 of 89.51, with competitive performance relative to published CT report generation baselines evaluated on the same split. On BIMCV-R, lexical overlap decreased substantially, whereas BERTScore F1 was 81.26.
Discussion:
This divergence indicates contextual embedding similarity despite substantial lexical and dataset shift; however, it does not establish factual, semantic, or clinical equivalence. The results demonstrate the retrospective computational feasibility of the proposed parameter-efficient design, while clinical accuracy and workflow utility were not prospectively evaluated.
Conclusion:
ReadCT provides a research-stage demonstration of compact, parameter-efficient findings generation from 3D chest CT volumes. Prospective workflow evaluation, structured factual assessment, and sufficiently powered multi-reader studies are required before clinical use can be considered.
