Bridging the Skill Gap in Clinical CBCT Interpretation with CBCTRepD¶

Conference: CVPR 2026 arXiv: 2603.10933 Code: N/A Area: Medical Imaging Keywords: CBCT report generation, oral and maxillofacial imaging, radiologist–AI collaboration, multi-level evaluation, bilingual reporting

TL;DR¶

This paper proposes CBCTRepD, a bilingual report generation system for oral and maxillofacial CBCT, trained on a high-quality paired dataset of 7,408 cases. A multi-level evaluation framework is introduced to validate its tiered empowerment effect on novice, intermediate, and senior radiologists within a radiologist–AI collaborative workflow.

Background & Motivation¶

Background: Oral and maxillofacial cone-beam CT (CBCT) is a widely used three-dimensional imaging modality in dental clinical practice. Generative AI has achieved notable progress in report generation for two-dimensional imaging modalities such as chest X-rays (e.g., R2Gen, CheXagent).

Limitations of Prior Work: CBCT report generation faces three major bottlenecks: (1) high-quality paired CBCT–report data are extremely scarce, with virtually no publicly available datasets; (2) CBCT is volumetric data involving 55 oral disease entities, making interpretation far more complex than two-dimensional imaging; (3) existing evaluation paradigms assess only the quality of direct AI output, overlooking the clinically common scenario of radiologist–AI collaboration.

Key Challenge: AI report generation technology is mature, yet the data infrastructure and evaluation frameworks lag behind — without large-scale paired data, reliable models cannot be trained; without collaborative evaluation frameworks, the actual clinical value of AI in real-world workflows cannot be measured.

Goal: To develop a CBCT report generation system that integrates into routine radiological workflows, and to quantify its empowerment effect on clinicians of varying experience levels through a multi-level evaluation framework.

Key Insight: Simultaneously addressing data, model, and evaluation dimensions — constructing a large-scale dataset, developing a bilingual report generation system, and designing a collaborative evaluation framework.

Core Idea: A CBCT report generation system is trained on 7,408 paired cases; an "AI draft + clinician editing" collaborative paradigm and a multi-level evaluation framework are employed to demonstrate clinically meaningful improvements across radiologists of all experience levels.

Method¶

Overall Architecture¶

CBCTRepD is an end-to-end report generation system for oral and maxillofacial CBCT. The pipeline proceeds as follows: CBCT volumetric data → 3D image encoder (extracting spatial features from multiple anatomical regions) → report generation module (automatically producing structured bilingual draft reports in Chinese and English) → collaborative editing interface (where clinicians revise and supplement the AI draft) → final clinical report. The system is designed as an AI collaboration tool embedded in routine radiological workflows.

Key Designs¶

Large-Scale High-Quality Paired Dataset:
Function: Construct the largest paired imaging–report dataset to date in the oral and maxillofacial CBCT domain.
Mechanism: Approximately 7,408 CBCT studies covering 55 oral disease entities (periodontitis, periapical lesions, jaw cysts, impacted teeth, temporomandibular joint disorders, etc.), encompassing multi-vendor devices, diverse scanning parameters, and varied clinical scenarios. Annotations are bilingual (Chinese and English) and reviewed by senior radiologists.
Design Motivation: High-quality paired data are the foundation for training reliable report generation models. Given the near-total absence of publicly available resources in this domain, the dataset itself constitutes a significant contribution.
Multi-Level Clinical Evaluation Framework:
Function: Establish an evaluation system covering two levels: "direct AI output" and "radiologist–AI collaborative output."
Mechanism: Layer 1 evaluates the quality of AI-generated report drafts independently; Layer 2 evaluates the quality of final reports produced by radiologists of different experience levels (novice/intermediate/senior) after editing the AI drafts. Evaluation dimensions include automatic metrics (BLEU, ROUGE, CIDEr, etc.), radiologist assessments (completeness, standardization, accuracy), and clinician assessments (adequacy of diagnostic information, decision-support value).
Design Motivation: Conventional evaluation measures only AI output quality, whereas real clinical practice operates in a collaborative mode. The multi-level framework more accurately reflects the system's value in actual deployment.
Bilingual Structured Report Generation:
Function: Generate structured reports with both Chinese and English output.
Mechanism: Based on features extracted by the 3D image encoder, the system automatically generates systematic reports covering each anatomical region, with emphasis on detecting co-existing lesions across regions.
Design Motivation: Bilingual output enables cross-regional deployment; structured output enhances report standardization and reduces omissions.

Loss & Training¶

The system is trained end-to-end on approximately 7,408 paired cases using a standard medical report generation paradigm (cross-entropy loss combined with reinforcement learning fine-tuning to optimize metrics such as CIDEr). Training data are reviewed by senior radiologists to ensure annotation quality.

Key Experimental Results¶

Main Results¶

Evaluation Subject	Report Structural Standardization	Lesion Description Completeness	Writing Standardization	Cross-Regional Lesion Coverage
Novice Radiologist	Low	Frequent omissions	Non-standard	Low
CBCTRepD AI Draft	Intermediate level	Comparable to intermediate	Near intermediate	Systematic coverage
Intermediate Radiologist	Moderate	Occasional omissions	Standardized	Moderate
Senior Radiologist	High	Rarely omitted	Highly standardized	High

AI–Clinician Collaborative Effect¶

Clinician Experience Level	Without AI Assistance	With CBCTRepD	Improvement
Novice	Novice level	Near intermediate level	Significant gains in completeness and standardization; fewer omissions
Intermediate	Intermediate level	Near senior level	More accurate lesion descriptions; increased cross-regional attention
Senior	Senior level	Senior level+	Fewer omission-related errors, including clinically significant missed findings

Key Findings¶

Reports directly generated by CBCTRepD are comparable to those of intermediate-level radiologists in writing quality and standardization.
Radiologist–AI collaboration yields consistent and clinically meaningful improvements across all three experience levels.
The most pronounced improvement is observed among novice clinicians (a full tier uplift); for senior clinicians, the value lies in a "safety net effect" — reducing fatigue-induced omissions.
The system is particularly effective in prompting clinicians to attend to co-existing lesions across anatomical regions.

Highlights & Insights¶

The dataset contribution itself is of major value: the 7,408-case paired dataset covering 55 disease entities fills a critical data gap in the oral and maxillofacial CBCT domain.
The evaluation paradigm represents an innovation: shifting from "evaluating AI output" to "evaluating AI–human collaborative output" more closely reflects real-world clinical deployment.
Tiered empowerment across experience levels: unlike most systems that benefit only novice users, CBCTRepD demonstrates clear value for senior clinicians as well (via omission reduction), which is critical for driving clinical adoption.
Focus on omissions rather than false positives: the system is especially effective at reducing omission errors, which carry greater safety implications than false positives in clinical radiology.

Limitations & Future Work¶

Data sources may exhibit single-center bias; multi-center validation is a prerequisite for generalization.
Sample distribution across 55 disease entities may be imbalanced, raising concerns about report generation quality for rare conditions.
The computational and storage overhead of CBCT volumetric data is substantial; inference efficiency and hardware requirements for real-world deployment warrant optimization.
The collaborative mode may introduce automation bias — awareness of AI assistance may alter clinicians' diagnostic behavior, and long-term effects require longitudinal investigation.
Direct quantitative comparison with other medical report generation methods (e.g., RadFM, LLaVA-Med) is absent.

vs. R2Gen / CheXagent: These methods have achieved success in report generation for two-dimensional chest X-ray imaging, but automatic report generation for three-dimensional CBCT was previously almost nonexistent; CBCTRepD fills this gap.
vs. RadFM / LLaVA-Med: Multimodal medical large language models have explored the potential of AI-assisted radiology, but lack systematic evaluation frameworks for collaborative scenarios; CBCTRepD's multi-level evaluation framework is an important complement.
vs. Dental Imaging AI (ToothNet, etc.): Existing dental AI primarily focuses on tooth segmentation and lesion detection, with little attention to automatic report generation.

Rating¶

Novelty: ⭐⭐⭐⭐ First report generation system targeting oral and maxillofacial CBCT; the dataset and evaluation framework fill a domain gap.
Experimental Thoroughness: ⭐⭐⭐⭐ The multi-level evaluation framework is well-designed, encompassing automatic metrics, radiologist expert assessment, and clinician evaluation.
Writing Quality: ⭐⭐⭐⭐ The problem–method–result logical chain is coherent, with high information density in the abstract.
Value: ⭐⭐⭐⭐⭐ Directly applicable to oral radiology clinical practice; the validated tiered empowerment effect serves as a model for AI deployment and adoption.