Preliminary study on AI-assisted diagnosis of bone remodeling in chronic maxillary sinusitis

Caiyun Zou, Hongbo Ji, Jie Cui, Bo Qian, Yu-Chen Chen, Qingxiang Zhang, Shuangba He, Yang Sui, Yang Bai, Yeming Zhong, Xu Zhang, Ting Ni, Zigang CheBmc Medical Imaging 2024 · DOI

TL;DR

This study introduces a two-stage AI approach using deep learning (CNN) and machine learning (SVM) to diagnose chronic maxillary sinusitis (CMS) and its associated bone remodeling from CT scans. The AI models achieved high accuracy, demonstrating the potential for standardized and efficient clinical diagnosis, addressing a critical need for better tools in managing this complex condition.

What Problem Does This Paper Address?

Chronic maxillary sinusitis (CMS) is a prevalent, long-term inflammatory disease of the nasal sinuses that can lead to significant bone changes, known as bone remodeling. This remodeling is a key factor in the persistence and recurrence of sinusitis, making treatment challenging and imposing a heavy burden on patients. Current diagnostic methods, while utilizing CT scans, often rely on subjective assessments like the Lund-Mackay scoring system, which can be time-consuming, inconsistent, and insensitive to subtle changes. There's a clear need for more objective, efficient, and accurate diagnostic tools to identify both CMS and its bone remodeling component. While AI has shown promise in medical imaging, its application to the combined diagnosis of CMS and bone remodeling, particularly using CT data, has been limited.

Methods

The researchers employed a two-stage AI strategy:

  • Stage 1: Deep Learning for CMS Detection: A Convolutional Neural Network (CNN) was developed for rapid binary classification of chronic maxillary sinusitis. This involved training the CNN on DICOM-format CT images, with regions of interest (ROIs) delineated by junior doctors and verified by senior radiologists. Image enhancement techniques were applied to the training data.
  • Stage 2: Machine Learning for Bone Remodeling Detection: A Support Vector Machine (SVM) model was used to identify bone remodeling in patients with CMS. This stage involved extracting imaging omics features from CT scans, followed by feature standardization and selection. The SVM model was trained on these features to discriminate between bone remodeling and its absence.

Both models were trained and tested on separate datasets, with performance evaluated using metrics like sensitivity, specificity, accuracy, and Area Under the Curve (AUC).

Datasets

  • Total Samples: 1000 sides from 500 patients.
  • Stage 1 (CMS Detection):
    • Training Set: 368 sides
    • Test Set: 93 sides
  • Stage 2 (Bone Remodeling Detection):
    • Training Set: 641 sides
    • Test Set: 161 sides

The data was collected from patients at Nanjing Tongren Hospital between January 2018 and December 2021. Exclusion criteria included fungal sinusitis, sinus tumors, cysts, cystic fibrosis, trauma, osteitis, and other non-CMS conditions with bone changes.

Key Insights and Ideas

  • Synergistic AI Approach: Combining a deep learning CNN for initial CMS detection with a machine learning SVM for bone remodeling offers a robust, two-pronged diagnostic strategy. This is particularly useful for handling imbalanced datasets and complex diagnoses.
  • CT Data Superiority: Utilizing 3D CT data for AI models provides a more objective and comprehensive assessment of the entire maxillary sinus cavity compared to 2D imaging or X-rays, which can suffer from artifacts and limited resolution.
  • Addressing Subjectivity: The AI models aim to reduce the subjectivity inherent in manual CT interpretation and scoring systems like the Lund-Mackay score, leading to more standardized and reproducible diagnoses.
  • Potential for Clinical Auxiliaries: The developed AI models can serve as valuable auxiliary tools for clinicians, potentially improving diagnostic efficiency and aiding in treatment planning.

Key Results

  • CMS Detection (CNN):
    • Test Set (93 samples): Sensitivity of 0.9796, Specificity of 0.8636, Accuracy of 0.9247, and AUC of 0.94.
  • Bone Remodeling Detection (SVM):
    • Test Set (161 samples): Sensitivity of 0.7353, Specificity of 0.9685, Accuracy of 0.9193, and AUC of 0.89.

The CNN model demonstrated excellent performance in identifying CMS, while the SVM model showed high specificity and good accuracy in detecting bone remodeling, even with a more challenging dataset.

Limitations and Future Work

  • Sample Size and Generalizability: The authors acknowledge that the CNN and SVM models may not be fully generalizable due to the limited sample size and optimization. They plan to accumulate more samples in future studies.
  • Subjective Measurement Bias: The definition of mucosal thickening (e.g., the 3 mm boundary) can introduce subjective measurement errors, potentially biasing the AI model's training.
  • Limited Scope of Bone Remodeling: The SVM model focused on the posterolateral wall of the maxillary sinus due to practical considerations of labeling large ROIs. Future work should expand the evaluation to other sinus areas, such as the sphenoid sinus, and investigate different clinical symptoms associated with bone remodeling in these regions.
  • Model Integration: While the study presents two separate models, future research could explore integrating them into a single, more comprehensive diagnostic system.
  • Further Diagnosis: The authors suggest considering further diagnosing other types of sinusitis in future studies.

Summary updated

BibTeX
@article{zou2024preliminary,
  author = {Caiyun Zou and Hongbo Ji and Jie Cui and Bo Qian and Yu-Chen Chen and Qingxiang Zhang and Shuangba He and Yang Sui and Yang Bai and Yeming Zhong and Xu Zhang and Ting Ni and Zigang Che},
  title = {{Preliminary study on AI-assisted diagnosis of bone remodeling in chronic maxillary sinusitis}},
  journal = {Bmc Medical Imaging},
  year = {2024},
  doi = {10.1186/s12880-024-01316-2},
}
Figure