📄 medicine

Development and Clinical Validation of an Enhanced Deep Learning Model for Automated Segmentation and Implant Planning in Maxillary Edentulous Regions Using CBCT Images

This study developed and clinically validated a deep learning framework, specifically a Swin-UNETR model, which achieves high-accuracy automated segmentation of maxillary edentulous regions in CBCT images, significantly reducing implant planning time and improving reproducibility compared to conventional workflows.

Original authors: Mehdi Abrishami, Hasan Momeni, Mehdi Rahmani, Elham Sheykhi Some

Published 2026-07-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mehdi Abrishami, Hasan Momeni, Mehdi Rahmani, Elham Sheykhi Some

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a master architect tasked with building a new house (a dental implant) on a plot of land that has been cleared of trees (a toothless upper jaw). To build safely, you need a perfect 3D map of the soil, knowing exactly where the soft dirt ends and the hard rock begins, and where the underground pipes (nerves and sinuses) are located.

In the past, architects (dentists) had to draw this map by hand, looking at a stack of 2D X-ray slices and mentally piecing them together. It was slow, tiring, and two different architects might draw slightly different maps for the same plot of land.

This paper describes a team of researchers who built a super-smart digital assistant to help draw that map automatically. Here is how they did it and what they found, explained simply:

The Problem: The "Hand-Drawn Map" Struggle

The researchers looked at 450 3D scans (CBCTs) of people's upper jaws who were missing some teeth. They wanted to see if a computer could automatically find the "empty land" (the edentulous ridge) where an implant could go.

The old way: Three expert dentists had to manually trace the boundaries of the bone on a computer for every single scan. It took a long time, and even experts sometimes disagreed on where the line should be drawn.
The challenge: The upper jaw is tricky. It has a bumpy floor (the sinus) right above it, and the bone can be very thin or irregular.

The Solution: Teaching the Computer to "See"

The team trained four different types of "digital brains" (Deep Learning models) to recognize these empty spaces. Think of these models as different students taking a test:

U-Net: The classic student who knows the basics.
Attention U-Net: A student who learned to focus on the most important details.
nnU-Net: A very organized student who follows a strict checklist.
Swin-UNETR: The "super-star" student. This model uses a special technology called a "Transformer" (the same kind of tech that helps computers understand human language). It's great at looking at the whole picture at once, not just small pieces, which helps it understand complex shapes like the upper jaw.

The Results: The "Super-Star" Wins

After training on hundreds of scans, they tested the models on new, unseen cases.

The Winner: The Swin-UNETR model was the clear champion. It matched the experts' hand-drawn maps with 92.1% accuracy.
The Comparison: The older models were good (around 84–90% accuracy), but the Transformer-based model was significantly better at handling the tricky, bumpy edges of the jawbone.
The "Mistakes": When the computer and the human experts disagreed, it was usually because the human experts had smoothed out the map too much or missed a tiny sliver of thin bone. In some cases, the computer actually drew a more accurate line than the human!

The Real-World Test: Speed and Consistency

The researchers didn't just stop at computer scores; they put the tool to the test in a real-world scenario. They asked 20 experienced dentists to plan implants on 20 different patients using two methods:

The Old Way: Drawing the map by hand.
The New Way: Using the AI to draw the map first, then just checking it.

The findings were impressive:

Time Saved: The AI cut the planning time almost in half. Instead of taking nearly 15 minutes per patient, it took about 8 minutes. That's a 43.5% time savings.
Better Agreement: When dentists used the AI, they all agreed much more on where to place the implant. Without AI, their plans varied more; with AI, they were on the same page.
Happy Users: The dentists loved the tool. They gave it high scores for usability and satisfaction, saying it made their job easier and less stressful.

The Bottom Line

This study shows that advanced AI, specifically the "Transformer" type (Swin-UNETR), can act like a highly skilled co-pilot for dentists. It doesn't replace the dentist, but it does the heavy lifting of drawing the map, allowing the dentist to focus on the final decision.

Important Note from the Paper:
The researchers are careful to say this is a validation study. They proved the AI works well for segmenting (drawing the boundaries of) the bone. They did not claim the AI automatically picks the final implant size or position on its own; the dentist still makes the final call. They also noted that while the results are great, the tool needs to be tested in more hospitals (multicenter studies) before it becomes a standard tool in every dental office.

In short: The AI is a fast, accurate, and consistent assistant that helps dentists draw the map of the jawbone much faster and with fewer disagreements, but the dentist is still the captain of the ship.

Technical Summary: Enhanced Deep Learning for Maxillary Edentulous Segmentation and Implant Planning

Problem Statement
Digital implant planning for partially edentulous maxillary regions relies heavily on Cone-Beam Computed Tomography (CBCT) imaging. However, the current workflow is operator-dependent, time-consuming, and susceptible to interobserver variability. Clinicians must manually delineate edentulous ridges, evaluate bone dimensions, and navigate complex anatomical structures like the maxillary sinus and nasal cavity. Existing artificial intelligence (AI) solutions have largely focused on tooth segmentation or mandibular structures, leaving a gap in robust, automated segmentation for the maxillary edentulous ridge, which presents unique challenges due to variable resorption patterns and complex morphology. Furthermore, previous studies have been limited by small datasets, heterogeneous annotation protocols, and a lack of clinical validation involving experienced practitioners.

Methodology
The study employed a diagnostic accuracy and clinical validation design using a retrospective dataset of 450 CBCT scans, from which 320 cases with partial maxillary edentulism were selected based on strict inclusion and exclusion criteria (e.g., residual ridge width ≥4 mm, absence of severe artifacts or pathology).

Data Preparation: Manual segmentation was performed by three board-certified oral and maxillofacial surgeons using 3D Slicer. A consensus annotation process, adjudicated by a fourth senior surgeon when necessary, generated the ground truth masks in NIfTI format.
Model Development: Four deep learning architectures were implemented and trained using the MONAI framework on PyTorch:
1. Conventional 3D U-Net
2. Attention U-Net
3. nnU-Net
4. Swin-UNETR (a transformer-based architecture)
  The dataset was split 80:10:10 for training, validation, and testing. Models utilized intensity normalization, resampling, and extensive data augmentation (rotations, flipping, elastic deformation). Training employed a combination of Dice and cross-entropy loss functions with early stopping.
Evaluation Metrics: Segmentation performance was quantified using the Dice Similarity Coefficient (DSC), Intersection over Union (IoU), Hausdorff Distance (HD95), Precision, Recall, and F1-score.
Clinical Validation: Twenty experienced clinicians performed implant planning on 20 CBCT scans under two conditions: conventional manual planning and AI-assisted planning (using the best-performing model). Metrics included planning time, number of manual corrections, interobserver agreement (ICC), System Usability Scale (SUS) scores, and user satisfaction.

Key Contributions

Architectural Comparison: The study provides a comparative analysis of CNN-based and Transformer-based architectures specifically for maxillary edentulous segmentation, identifying Swin-UNETR as the superior model.
Clinical Workflow Integration: Unlike purely technical studies, this research integrates the AI model into a clinical validation phase, measuring its impact on planning time, consistency, and usability among practicing clinicians.
Discrepancy Analysis: The paper offers a detailed analysis of failure modes, distinguishing between model errors and limitations inherent in manual annotation (e.g., smoothing artifacts, omission of thin bone).

Results

Segmentation Performance: All models successfully segmented the target regions. Swin-UNETR achieved the highest performance with a mean DSC of 92.1%, followed by nnU-Net (90.4%), Attention U-Net (87.8%), and U-Net (84.3%). Swin-UNETR also demonstrated the lowest Hausdorff Distance (1.8 mm).
Clinical Efficiency: AI-assisted planning reduced the mean planning time from 14.7 minutes to 8.3 minutes, a 43.5% reduction ( $p < 0.001$ ).
Consistency and Usability: The use of AI significantly improved interobserver agreement regarding implant position and angulation (ICC increased from 0.84 to 0.93). The number of manual corrections required dropped from 3.6 to 1.2 per case. System Usability Scale (SUS) scores were excellent (88.6), and user satisfaction was high (4.7/5).
Discrepancies: In cases where the AI and manual segmentation differed, the AI often provided masks that more accurately followed anatomical contours (e.g., sinus floor) than the consensus manual labels. Common sources of disagreement included manual smoothing procedures and the exclusion of thin residual bone by human annotators.

Significance and Claims
The authors claim that the proposed deep learning framework, particularly the transformer-based Swin-UNETR architecture, provides accurate and reliable segmentation of maxillary edentulous regions, outperforming conventional CNNs in capturing complex anatomical structures. The study posits that integrating AI into digital implant planning workflows significantly enhances efficiency, reproducibility, and standardization, thereby addressing the variability associated with operator experience.

However, the authors maintain a modest stance regarding clinical translation. They acknowledge that the study was retrospective and single-centered, necessitating further multicenter validation to confirm generalizability. Additionally, they clarify that the current model is designed for segmentation and does not yet automatically determine final implant dimensions or positions, identifying the need for future research into end-to-end, fully automated prosthetically driven planning systems.