ENTERPRISE AI ANALYSIS
Automated detection and classification of maxillary sinus variations using slice-based and full-volume CBCT deep learning models
Zainab Abdulkader Said, Abbas Ahmed Abdulqader, Fayu Liu, Fu Lin Jiang, Juan Li, Chong Zhang, Fangyuan Cheng, Thekra Ali Saeed & Baleegh Abdulraoof Alkadasi
This study aimed to develop and compare two deep learning models for automated detection and classification of maxillary sinus variations using cone-beam computed tomography (CBCT): a slice-based two-dimensional (2D) model based on sagittal images and a three-dimensional (3D) volume model using full CBCT volumetric scans. CBCT scans from 452 patients (631 sinuses) were reviewed and categorized into six clinically relevant sinus radiographic variations: normal anatomy, hypoplasia, mucosal thickening, polypoid lesions, septa, and sinus opacification. For the two-dimensional slice-based model, 7,232 sagittal images were initially extracted; after quality screening, 1,880 representative slices were selected for model development. Three convolutional neural network architectures were evaluated, with DenseNet-121 demonstrating the best performance. For the 3D model, all sinuses were manually annotated using 3D Slicer to define inner and outer sinus regions of interest. Both models were independently trained and evaluated using standard classification performance metrics, including sensitivity, specificity, precision, F1 score, and area under the receiver operating characteristic curve. The 2D slice-based evaluation model achieved an overall accuracy of 83.2%, with high sensitivity for septa (0.93), normal anatomy (0.88), and polypoid lesions (0.81); however, it had lower sensitivity for hypoplasia (0.67). The 3D volume-based model demonstrated superior performance, achieving an accuracy of 87.2%, with improved sensitivity for hypoplasia (0.88), mucosal thickening (0.93), and polypoid lesions (0.75), as well as perfect classification of sinus opacification with both sensitivity and specificity: 1.00. Both slice-based and volume-based deep learning models showed strong potential for automated classification of maxillary sinus variations on CBCT images. While the 2D slice-based model offers a fast and computationally efficient approach, the full-volume 3D model benefits from enhanced spatial representation and higher diagnostic precision. These results highlight the potential of Artificial Intelligence as an adjunctive tool in radiographic assessment of the maxillary sinus.
Revolutionizing Maxillary Sinus Diagnostics with AI
This research pioneers deep learning for automated detection and classification of maxillary sinus variations, leveraging both 2D slice-based and 3D full-volume CBCT models. The findings demonstrate significant improvements in diagnostic accuracy and efficiency, offering a transformative tool for dental implantology and surgical planning.
0
Overall Accuracy (3D Model)
0
Overall Accuracy (2D Model)
0
Sinus Opacification Sensitivity (3D)
0
Septa Sensitivity (2D)
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Study Objective
This study aimed to develop a deep learning framework for the automated diagnosis of maxillary sinus variations, encompassing six sinus radiographic variation features using both 2D sagittal images and 3D full volume data. The goal was to enhance diagnostic accuracy, minimize observer variability, and optimize treatment planning in the posterior maxillary region.
Data Processing Workflow
CBCT Scan Acquisition (452 Patients)
→
Sinus Categorization (631 Sinuses)
→
2D Slice & 3D Volume Extraction
→
Manual Annotation (3D Slicer)
→
Deep Learning Model Training
→
Performance Evaluation
Data Collection & Preprocessing
A retrospective cohort study included 452 patients (631 maxillary sinuses) analyzed using Invivo software. Patients ranged from 14 to 90 years. CBCT scans were obtained using a dental cone-beam system with specific parameters, reconstructed to 536 × 536 × 448 voxels at 0.3 mm isotropic resolution, and stored with intensity values from -1000 to 7548 Hounsfield Units.
83.2%
The 2D slice-based evaluation model achieved a strong overall accuracy, demonstrating its utility for rapid initial screening.
2D Model Performance Metrics (DenseNet-121)
| MS Variation | Sensitivity | Specificity | Precision | F1 Score | AUC |
|---|---|---|---|---|---|
| Hypoplasia | 0.67 | 0.97 | 0.64 | 0.65 | 0.83 |
| Mucosal Thickening | 0.88 | 0.95 | 0.83 | 0.85 | 0.91 |
| Normal | 0.88 | 0.99 | 0.98 | 0.92 | 0.94 |
| Polypoid Lesions | 0.81 | 0.96 | 0.79 | 0.8 | 0.89 |
| Septa | 0.92 | 0.98 | 0.93 | 0.93 | 0.96 |
| Sinus Opacification | 0.83 | 0.98 | 0.71 | 0.77 | 0.91 |
Note: DenseNet-121 showed the highest sensitivity (0.86) among tested architectures for 2D slice-based classification.
Model Architecture and Training (2D)
For the 2D slice-based model, 1,880 representative sagittal slices were selected. Three CNN architectures (ResNet-34, ResNet-101, DenseNet-121) were evaluated, with DenseNet-121 performing best. Images were resized to 416x416 pixels. Training used an 80/20 split, Adam optimizer (learning rate 0.001), batch size 32, ReLU activation, and cross-entropy loss. Data augmentation included rotation and Gaussian noise. Training ran for 169 epochs on an NVIDIA GeForce RTX 3090 GPU (14.2h).
87.2%
The 3D volume-based model demonstrated superior performance, providing enhanced spatial representation and higher diagnostic precision.
3D Model Performance Metrics (DenseNet-121)
| MS Variation | Sensitivity | Specificity | Precision | F1 Score | AUC |
|---|---|---|---|---|---|
| Hypoplasia | 0.88 | 0.97 | 0.84 | 0.86 | 0.92 |
| Mucosal Thickening | 0.93 | 0.92 | 0.70 | 0.80 | 0.92 |
| Normal | 0.75 | 1.00 | 1.00 | 0.86 | 0.88 |
| Polypoid Lesion | 0.75 | 0.98 | 0.86 | 0.80 | 0.87 |
| Septa | 0.92 | 0.96 | 0.83 | 0.87 | 0.94 |
| Sinus Opacification | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 |
Note: Perfect classification for sinus opacification was achieved by the 3D model.
Advantages of 3D Volumetric Data
The 3D volume-based model leveraged full spatial information and ROI annotations to deliver a more detailed representation of sinus anatomy. This approach allowed for better localization and differentiation of subtle or overlapping abnormalities that may be missed in 2D slices. Specifically, it showed a higher TPR for hypoplasia (0.88 vs. 0.67) and perfect classification for sinus opacification (TPR = 1.00 vs. 0.83), underscoring the value of 3D spatial context.
Strengths of Each Model
While both models show strong potential, the 2D slice-based model offers a fast and computationally efficient approach suitable for initial screening and detecting common features like mucosal thickening and septa. The 3D full-volume model, however, excels in identifying more complex conditions, including hypoplasia and complete sinus opacification, benefiting from enhanced spatial representation and higher diagnostic precision. This dual-modality design, applied to a large and diverse dataset, enhances generalizability and robustness.
2D vs. 3D Model Performance Highlights
| Feature | 2D Slice-Based Model | 3D Volume-Based Model |
|---|---|---|
| Overall Accuracy | 83.2% | 87.2% |
| Hypoplasia Sensitivity | 0.67 | 0.88 |
| Sinus Opacification Sensitivity | 0.83 | 1.00 |
| Septa Detection | Good (TPR 0.92) | Good (TPR 0.92) |
| Speed & Efficiency | High | Moderate |
| Spatial Context | Limited | Comprehensive |
Note: The 3D model generally outperformed the 2D model for complex conditions, while 2D offered speed.
Clinical Implications and Future Potential
The automated and reliable classification of maxillary sinus variations has significant implications for enhancing treatment outcomes. Accurate detection of sinus septa is critical for minimizing Schneiderian membrane perforation during SFE procedures. Our model can serve as a clinical decision-support tool, enhancing diagnostic accuracy, streamlining preoperative planning, and improving risk stratification, ultimately supporting personalized care. Future work should incorporate larger, multi-center datasets and dedicated detection strategies for class imbalance.
Calculate Your Potential ROI
Estimate the financial and operational benefits of implementing AI solutions in your enterprise.
Estimated Annual Savings
$0
Hours Reclaimed Annually
0
Your AI Implementation Roadmap
A typical phased approach to integrate advanced AI diagnostics into your practice.
Data Ingestion & Preprocessing
Gather and clean CBCT data, standardize formats, and perform initial augmentations for model training.
Model Architecture Selection & Training
Evaluate and train optimal deep learning models (DenseNet-121) using both 2D slice-based and 3D volumetric data.
Validation & Performance Tuning
Rigorously test models against unseen data, fine-tune parameters, and validate accuracy, sensitivity, and specificity.
Clinical Integration & Deployment
Integrate the validated AI models into existing clinical workflows, ensuring seamless operation and user adoption.
Continuous Learning & Monitoring
Implement feedback loops for ongoing model improvement and monitor performance in real-world clinical settings.
Ready to Transform Your Diagnostic Capabilities?
Book a free consultation with our AI specialists to explore how these advanced deep learning solutions can be tailored for your enterprise.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat