Generating Findings for Jaw Cysts in Dental Panoramic Radiographs Using a GPT-Based VLM: A Preliminary Study on Building a Two-Stage Self-Correction Loop with Structured Output (SLSO) Framework

Technical Deep Dive

Overview

This study explored using a GPT-based vision-language model (VLM) to generate radiological findings for jaw cysts in dental panoramic radiographs. The key contributions were:

• Proposed an integrated "Self-correction Loop with Structured Output (SLSO)" framework to improve the consistency and reliability of AI-generated findings. The framework incorporated:
  • Structured data generation to reduce ambiguity
  • Iterative self-correction loops to ensure consistency
  • Explicit negative finding descriptions and hallucination suppression
• Compared the SLSO framework against a conventional Chain-of-Thought (CoT) approach, finding:
  • Improved accuracy in tooth number identification (+66.9% relative), root resorption, and anatomical relationship descriptions
  • Enforcement of comprehensive findings, including explicit negative details
  • Reduction in hallucinations and logically inconsistent statements

Methodology

• Dataset: 22 dental panoramic radiographs with annotated jaw cysts, collected and labeled by a dental radiologist
• Implemented a 10-step SLSO framework incorporating:
  • Image analysis using GPT-4o
  • Structured data generation with a predefined schema
  • Tooth number extraction and consistency checks
  • Natural language finding generation and reverification

Results

• The SLSO framework showed consistent improvements over the CoT baseline, especially in:
  • Tooth number identification accuracy (66.9% relative increase)
  • Descriptions of root resorption, tooth displacement, and anatomical relationships
• Successful cases demonstrated concise, clinically appropriate findings after 5 regeneration cycles on average
• Limitations remained in accurately describing extensive lesions spanning multiple teeth and identifying subtle changes like mild resorption

Interpretation

• The SLSO framework functioned as an external validation mechanism, enhancing the reliability of VLM-generated radiological findings
• Key effects included:
  • Improved precision in localization through structured tooth number tracking
  • Comprehensive documentation of negative findings
  • Suppression of hallucinations and logically inconsistent statements
• The framework represents a model-agnostic technical strategy, complementing model-level hallucination mitigation approaches

Limitations & Uncertainties

• Small dataset size (22 cases) limits statistical power and generalizability
• Challenges in accurately interpreting complex anatomical relationships and subtle changes
• Difficulty in directly comparing structured output generation due to model safety constraints

Future Work

• Validation on larger, more diverse datasets with multi-rater consensus ground truth
• Incorporation of semantic similarity metrics and expert-based rating scales to better capture clinical utility
• Exploration of the framework's applicability to other dental pathologies and medical imaging domains