Design and implementation of a radiomic-driven intelligent dental hospital diversion system utilizing multilabel imaging data

Yanchan Wu, Changyuan Yu, Meijia Zhang, Yichen Li, Yijun Wang, Dongren Yang, Yun Yang, Hao Lou, Chufan Ren, Enna Cai, Chenyue Dai, Ryan Sun, Qiang Xu, Qi Zhao, Huanhuan Zhang, Jiefan LiuJournal of Translational Medicine 2024 · DOI

Abstract

The main contribution of this study is to propose a new treatment process that incorporates biological multi-omics data, which reduces the workload of physicians while providing timely and accurate medical care to patients. Through comparative experiments, we demonstrate that the process is more efficient than existing processes. In addition, this intelligent triage system demonstrates high prediction accuracy in practical applications, providing new ideas and methods for biological multi-omics research.

Introduction

The global burden of oral disease is increasing, yet dental triage remains a manual, subjective, and often inefficient process. In high-volume clinical settings, patients frequently struggle to identify the appropriate specialist department, such as orthodontics versus endodontics. Traditional automated triage systems often rely on patient-provided text or voice descriptions, which are susceptible to subjective bias and misinterpretation. This leads to resource misallocation, extended wait times, and increased clinician burnout. This study addresses the "triage bottleneck" by implementing an objective, radiomic-driven system that utilizes X-ray imaging data for hospital diversion.

Methodology

Datasets and Preprocessing

The study utilized a substantial dataset collected between July and August 2023. The initial pool consisted of 79,139 case records and 4,682 X-ray data points. After rigorous cleaning to remove unidentified, unlabeled, or non-dental related cases, a final dataset of 3,943 unique dental X-rays was established.

Data selection and preprocessing flowchart

The taxonomy included six major departments: Oral Implantology, Prosthodontics, Periodontology, Orthodontics, Maxillofacial Surgery, and Endodontics, covering 77 specific conditions. The data was partitioned into development (n=2,523), validation (n=788), and external optimization (n=632) cohorts.

Model Architecture

The researchers developed a multi-label learning framework capable of testing several deep learning architectures, including DenseNet121, ResNet50, EfficientNet-B7, and MobileNet-V2. Unlike standard single-label classification, this system accounts for overlapping dental issues visible in a single X-ray.

Radiomic-driven intelligent diversion system architecture

Key technical strategies included:

  • Binary Cross Entropy (BCE) Loss: Utilized with an unbalanced weighting strategy to address rare dental conditions.
  • Data Augmentation: The training set was expanded three-fold using rotation, horizontal flipping, and contrast enhancement to improve robustness.
  • Deployment: The optimized model was integrated into an Android application for real-time clinical use.

Results

DenseNet121 emerged as the superior architecture, achieving a classification accuracy (ACC) of 0.80 in the test cohort and 0.82 in the external validation cohort. The "dense connections" within the model were particularly effective for medical images where feature reuse is critical.

Clinical trials demonstrated significant efficiency gains. The automated system saved patients an average of one hour per visit by bypassing the preliminary generalist consultation. Comparative analysis showed that DenseNet121 and ResNet50 outperformed both lightweight (MobileNet-V2) and overly complex (EfficientNet-B7) models.

Discussion and Conclusion

A critical insight from the study was the initial "tooth-centric" bias of the AI, which struggled with periodontology and maxillofacial diseases by focusing primarily on teeth rather than surrounding structures. This underscores the need for improved spatial awareness in medical AI.

While the system shows great promise, limitations include the narrow scope of a single-center dataset and the need for broader regional validation. Future work will focus on multi-modal integration—combining X-rays with Electronic Medical Records (EMR)—to create a more holistic "biological multi-omics" triage solution. This radiomic-driven approach provides a scalable solution for improving the efficiency and accuracy of dental triage.

Summary updated

Abstract

The main contribution of this study is to propose a new treatment process that incorporates biological multi-omics data, which reduces the workload of physicians while providing timely and accurate medical care to patients. Through comparative experiments, we demonstrate that the process is more efficient than existing processes. In addition, this intelligent triage system demonstrates high prediction accuracy in practical applications, providing new ideas and methods for biological multi-omics research.
BibTeX
@article{wu2024design,
  author = {Yanchan Wu and Changyuan Yu and Meijia Zhang and Yichen Li and Yijun Wang and Dongren Yang and Yun Yang and Hao Lou and Chufan Ren and Enna Cai and Chenyue Dai and Ryan Sun and Qiang Xu and Qi Zhao and Huanhuan Zhang and Jiefan Liu},
  title = {{Design and implementation of a radiomic-driven intelligent dental hospital diversion system utilizing multilabel imaging data}},
  journal = {Journal of Translational Medicine},
  year = {2024},
  doi = {10.1186/s12967-024-05958-2},
}
Figure