Estimation and angular analysis of anatomical reference points on cephalometric films with artificial intelligence approach

A substantial corpus of research on AI applications has demonstrated that the quality and efficiency of healthcare services can be enhanced, while concomitantly reducing costs and human/system errors. The advantages of AI in healthcare are manifold, including but not limited to the following: time saving, standardisation through automation of tasks requiring manual skills, and reduced workload. The utilisation of artificial intelligence (AI) components has become increasingly prevalent in the domains of orthodontics and jaw surgery, with the aim of enhancing the efficacy of treatment and diagnostic procedures. In the domain of orthodontics, particularly in the realm of orthognathic surgery, the utilisation of cephalometric films assumes a pivotal role in ensuring the efficacy and precision of treatment outcomes. In the context of contemporary discourse on AI, treatment methodologies are devised through the meticulous delineation of points and angles on cephalometric films. Nevertheless, challenges have been identified in the field of image analysis in healthcare. These challenges are the lack of sufficient medical expertise among the engineering team developing the algorithm and the lack of complete knowledge of the radiologists who will use the algorithm. It is evident that there is a necessity for more efficient and clearer programs, owing to the deficiencies and reliability issues that have been identified in the existing systems. Errors may be caused by incorrect use and interpretation as a result of these deficiencies. The present study was meticulously planned through a collaborative endeavour between engineers and physicians, with the objective of addressing these challenges and enhancing the efficacy of image analysis in the healthcare sector. Accurate estimation of anatomical points is vital for clinical evaluations and treatment decisions, and inaccurate manual analysis may cause inconsistencies in the measurement of various craniofacial parameters. Consequently, there is a necessity to develop an automatic anatomy point detection system that can accurately, reliably and rapidly determine anatomy points. For this reason, the current study developed a FARNet-based model to create a local artificial intelligence model. This model facilitates accurate orthodontic treatment planning by automatically detecting cephalometric anatomy points and providing measurements of skeletal angle values used in jaw surgeries. To this end, the current study used a dataset consisting of 902 lateral cephalogram films covering 46 anatomical points and spanning different age groups. This dataset was downloaded from the American Association of Orthodontists Foundation's open-access, internet-based Craniofacial Growth Legacy Collection and from https://github.com/JuvenileInWind/FARNet, as reported in Ao and Wu (2023). A labelled dataset was created by obtaining the coordinates of all anatomy points in 902 films: 721 for training, 80 for validation and 101 for testing. The FARNet network was utilised for the automatic detection of cephalometric anatomy points. The performance of the model was evaluated using the mean radial error and success detection rate (SDR) within the 2 mm, 2.5 mm, 3 mm and 4 mm ranges. The results of the detection of skeletal, dental and soft tissue points in two-dimensional cephalometric radiographs obtained from the developed model are encouraging. In the FARNet model, the study's architecture yielded SDR values above 80%, with the exception of three points (Glabella (G'), Pterygoid (Pt) and Ramus (Ra)) in the 2 mm range. The model's performance values in the study were as follows: training accuracy 0,977, validation accuracy 0,930, test accuracy 0,935, training loss 0,090, validation loss 0,170 and test loss 0,165. There was no overfitting in the model. The model showed good learning and there was no overfitting. Additionally, a skeletal angle measurement model was created using the outputs obtained from the FARNet architecture. The angular parameter values obtained from this model revealed consistent results. These findings suggest that the model has the potential to function as a decision support mechanism, thereby reducing the time burden on physicians' clinical routines. Future iterations should include dental and soft tissue angular and linear parameters in the study framework, particularly in the context of surgical and orthodontic planning and diagnostic studies. This study is expected to contribute to the advancement of clinical knowledge and expertise by increasing its comprehensiveness and practical utility. In addition, the study presents a comparison of the patient's WebCeph output with the output of the developed artificial intelligence program and an evaluation conducted with an expert maxillofacial surgeon with 20 years of experience in the field. The results of this study show that the developed artificial intelligence model holds great promise for cephalometric analysis and the evaluation of preoperative movement amounts, which form a basis for diagnosis, treatment planning and follow-up in clinical orthodontic practice.