| Abstract | Svrha ovog rada je razvijanje i testiranje modela umjetne inteligencije u otkrivanju radiolucentnih lezija donje čeljusti. Ciste i tumori čeljusti učestali su klinički entiteti u svakodnevnom radu oralnog i maksilofacijalnog kirurga te se često otkrivaju sasvim slučajno tijekom radiološkog snimanja čeljusti. U početnim stadijima ne uzrokuju simptome što dovodi do kasnog dijagnosticiranja, a samim time dolazi do otežanog liječenja i potencijalno lošijeg ishoda za pacijente. Umjetna inteligencija je izvrstan alat u rješavanju problema brzog dijagnosticiranja lezija čeljusti. Duboko učenje, kao jedan dio umjetne inteligencije, koristi algoritme inspirirane strukturom i funkcijom ljudskog mozga, poznate kao umjetne neuronske mreže i iznimno su korisne u području medicinske dijagnostike. Istraživanje je obuhvatilo analizu 226 radiolucentnih lezija donje čeljusti dokumentiranih na
ortopantomogramima s potvrđenim patohistološkim dijagnozama. Uključene lezije su: radikularna cista, folikularna cista, ameloblastom, odontogena keratocista i rezidualna cista. Kako bi se povećala pouzdanost razvijenog modela umjetne inteligencije, uzorak je dodatno obogaćen korištenjem različitih metoda augmentacije slika. Točnost razvijenog modela umjetne inteligencije analizirana je putem metrika: preciznost, osjetljivost, srednja prosječna preciznost i krivuljom Preciznost-Osjetljivost.
U zadatku detekcije, preciznost, osjetljivost i prosječna preciznost mAP@50 su: 92.5%, 81% i 95.2%. Slično, u zadatku segmentacije, razvijeni model imao je preciznost, osjetljivost i prosječnu preciznost: mAP@50 od 100%, 94.5% i 96.7%. Tijekom testiranja razvijenog modela u kombiniranom zadatku detekcije i dijagnostike, odnosno segmentacije i dijagnostike, došlo je do pada u metričkim rezultatima što je povezano s kompleksnosti zadatka. Rezultati ovog istraživanja pokazali su metrički značajne mogućnosti razvijenog modela u detekciji, segmentaciji i dijagnosticiranju radiolucentnih lezija donje čeljusti, kao i njegovu
buduću upotrebu u kliničkoj praksi. |
| Abstract (english) | Introduction: Jaw cysts and tumors are common clinical entities encountered in the daily practice of oral and maxillofacial surgeons, as well as dentists. Typically, they remain asymptomatic until their growth begins to alter the shape of surrounding anatomical structures or inflammation develops. As a result, they are often discovered incidentally during radiological examinations, leading to delayed diagnoses and, consequently, less favorable treatment outcomes. For oral and maxillofacial surgeons, early detection is crucial for devising an appropriate treatment plan and ensuring the best possible patient outcomes. Deep learning, a subset of artificial intelligence, employs algorithms inspired by the structure and function of the human brain, known as artificial neural networks. Comprising layers of nodes or "neurons," each layer performs specific computations on input data, proving to be immensely valuable in the medical field. These networks can be utilized for the rapid diagnosis of radiolucent jaw lesions, enhancing patient care and treatment efficiency.
Aim: The aim of this research is to design and evaluate an advanced aritficial intelligence model capable of automatically detecting and segmenting radiolucent lesions of the lower jaw, utilizing the cutting-edge You Only Look Once (YOLO) version 8 technology.
Materials and methods: In our study, we analyzed a collection of 226 panoramic radiographs, spanning from 2013 to 2023, sourced from the Clinical Hospital Dubrava and the School of Dental Medicine at the University of Zagreb. The panoramic radiographs in this study showcased a diverse array of radiolucent lesions, such as radicular cysts, ameloblastomas, odontogenic keratocysts (OKCs), dentigerous cysts, and residual cysts. Each lesion underwent verification through pathohistological analysis and received annotations within the GIMP software by a collaboration of radiologists and oral and maxillofacial surgeons. The development and evaluation of the artificial intelligence model were conducted at the Faculty of Electrical Engineering and Computing, University of Zagreb. To enrich our database and improve the model's performance, we enhanced the dataset through a series of image augmentation techniques, such as translation, scaling, rotation, horizontal flipping, and the application of mosaic effects. To accomplish our objectives in detection, segmentation, and diagnosis, we utilized the sophisticated features of the YOLOv8 deep neural network, harnessing its advanced capabilities for precise and efficient analysis.
Results: In the detection task, the precision, sensitivity, and mean Average Precision (mAP@50) were 92.5%, 81%, and 95.2%, respectively. Similarly, in the segmentation task, the developed model achieved a precision, sensitivity, and mean Average Precision (mAP@50) of 100%, 94.5%, and 96.7%. During the testing of the developed model on combined tasks
of detection and diagnosis, as well as segmentation and diagnosis, there was a decrease in metric results, which is attributed to the complexity of the tasks. Specifically, in the detection and diagnosis task, the precision, sensitivity, and mean Average Precision in the augmented set were 55.6%, 79.8%, and 73.1%, while the metric data in the segmentation
and diagnosis task were 64%, 66.2%, and 74.9%.
Conclusion: Artificial intelligence can be used in the detection, segmentation, and diagnosis of radiolucent lesions of the lower jaw. A model developed based on YOLOv8 has the ability to quickly diagnose cysts and tumors of the jaw. To our knowledge, this is the first study that has unified detection, segmentation, and diagnosis of lower jaw lesions using an image resolution of 2776x1480 pixels. |
