The Publishing Journey of Two Papers: From Panoramic X-rays to Deep Learning

I've redesigned my website countless times, but I've been meaning to share the story behind these two articles for a while now.

Before I begin, I want to express my heartfelt gratitude to my dear sister, whose contributions—from the data collection phase to the final publication—have been invaluable to both me and the academic community. Without her guidance and patience, this journey wouldn't have been nearly as meaningful.

On the surface, there are just two publications. But getting them to this point took quite a while, and the story is a bit long. I've broken down this journey into six short chapters below.

It all started with Soru Kayısı, a project I designed while in Malatya to help undergrads play Kahoot-style quizzes professionally and without any interruptions. I'm originally from Trabzon, but thanks to my love for Malatya and the connections I made there, the journey didn't stop with just that project.

We started off with the question, "How can we make this content more useful?" We began by acquiring the dataset, and what followed was a process that took nearly a year.

As the number of labels grew with each new batch of data, we were simultaneously trying to improve the model and strike the perfect balance between performance and stability. Setting aside the false positives, there were such tiny lesions and cavities in the panoramic X-rays that the faint, ambiguous areas left both the clinicians and me feeling uncertain.

At this point, inter-observer agreement methods proved to be quite effective, and eventually, Roboflow's auto detection feature came to our rescue. Taking advantage of this tech blessing really sped things up.

Since the lesion post was my first official publication, I was pretty hyped. Up until that point, I hadn't even used RunPod; I didn't even know which GPU was better suited for which task.

At this stage, I got some support from F. Gözükara, a lecturer at my university with extensive experience in AI modeling. Thanks to his guidance, I learned the ropes of the platform. I used the Runpod PyTorch 2.8.0 image on Ubuntu 22.04; I started running training sessions using an H100 on one and RTX PRO 6000 GPUs on the other.

In the beginning, we tested small models on Roboflow to find the most successful augmentation configurations. Then, I exported those datasets and moved on to the actual training on remote servers.

At first, waiting around late at night to monitor the training was a grind. But once the success rates started climbing and the results became satisfying, I realized this work actually requires the spirit of an explorer. We weren't exactly discovering America, but for a fresh computer engineering grad, working alongside academics with years of experience in the field was a rare opportunity.

I treated both of these projects as my top priority, just like my own work. I didn't care about the clock; we kept at it until I could finally say, "Alright, this feels right."

The first publication focuses on the automated detection of periapical lesions in teeth with fixed prostheses using panoramic radiographs, and the algorithmic analysis of the relationship between the lesion below the root and the crown or bridge above it.

In this study, we retrospectively selected 404 panoramic radiographs from the Inönü University Faculty of Dentistry and made a total of 1,686 annotations, including 1,033 crowns and 653 lesions. I trained five different YOLO11 segmentation variants over 150 epochs. The best result came from YOLO11l-seg: mAP50 0.885, recall 0.853, precision 0.847.

The most interesting results came from the crown-lesion relationship analysis. Approximately 85 percent of the lesions were associated with a crown. The incidence of lesions in mandibular crowns was 2.7 times higher than in maxillary crowns. The Python algorithm I developed was able to perform this matching with 95 percent accuracy.

The study was published in The Journal of Prosthetic Dentistry (Elsevier).

In our second publication, we focused on detecting open-margin caries developing under crowns to simplify the diagnostic process for clinicians.

We had the chance to apply what we learned from our first paper much more efficiently here. We scanned 1,742 panoramic X-rays, extracted 257 high-quality ROIs, and designed a two-stage coarse-to-fine strategy.

In the first stage, we localized the crowns using YOLO11l (mAP50 0.977). In the second stage, we compared three different architectures: YOLO11-seg variants, U-Net with a ResNet34 backbone, and U-Net++ with the same backbone. In pixel-based evaluation, U-Net++ stood out with a 71.2% Dice score. On the independent test set, we reached a 68.6% Dice score.

In this study, we also shared an interesting finding highlighting that model capacity should be chosen based on dataset size: we observed that performance dropped as the model size increased in YOLO11 variants. We presented all the results and evaluations to the readers of the European Journal of Therapeutics.

The short summary of this long post is this: continuing to research and study is the greatest personal pleasure for me. What keeps motivation high is always staying on the hunt. Somewhere, somehow, there is always a hidden door waiting to open for you.

That was how my first experience went. I hope yours turns out even better.

Publication Links

Detection of periapical lesions in teeth with fixed prostheses using segmentation models and analyzing crown–lesion relationships — The Journal of Prosthetic Dentistry (Elsevier), 2025.

A Two-Stage Deep Learning Approach Using YOLO11 and U-Net for Detection and Segmentation of Caries Under Crowns in Panoramic Radiographs — European Journal of Therapeutics, 2026.

1 / 2