Since the introduction of cephalometry in 1931 by Broadbent[1], it has become an important tool in clinical diagnosis, treatment planning, evaluation of treatment changes and growth study. Cephalometric analysis of linear and angular measurements between landmarks and superimposition analysis of series of cephalometric images are two main methods of cephalometry. The traditional method is hand tracing the craniofacial soft and hard anatomic structural contours on cephalograms on acetate paper. The process is subjective, and the accuracy varies with personal experience, knowledge, understanding and tracking preference[2–6]. Additionally, this process is time-consuming, and the accuracy is also inevitably influenced by the fatigue level of humans[7, 8]. In particular, for the purpose of research, a certain number of cephalograms need to be traced and measured within a certain time constraint, and the intra-/inter-reproducibility is impacted.
With the development of digital technology, traditional hand tracing cephalogram is being replaced by digital cephalometric analysis. In previous studies, the latter method using commercial software has been proven to be accurate, reliable and time saving[8–12]. However, the only exception is the structural superimposition for treatment evaluation and growth study, although much effort has been made in this field[3, 12–14]. The reason for this exception lies in the fact that landmark identification is essential for diagnosis purposes and is easy to fulfil using commercial software. In contrast, structural superimposition focuses on tracing of structural details, which is independent of landmarks, and is still not feasible using current commercial software.
Baumrind et al. claimed that magnification, tracing, landmark identification and measurements are major sources of error for cephalometric analysis [2, 15, 16] They believed that hand superimpositions on reference planes results in better quality than any computer-aided superimpositions because biological craniofacial growth is difficult to interpret by any mathematical equations[2]. However, due to its absolute consistency, fully automated cephalometric analysis has always been a popular challenge in computer science. One of these methods is the knowledge-based line extraction technique[17], which duplicates the strategy of orthodontists in which important anatomic edges are extracted and landmarks are located according to geometric definitions. However, the irregular details of bone, such as the inter-trabeculae, incisor nerve canal and inferior alveolar canal, make computer automated tracing difficult and questionable. Other studies have attempted to locate landmarks directly[18, 19], and the techniques have evolved from template matching [18] to, more recently, neural network models[19]. In terms of structural superimposition on stable regions instead of reference planes, which has been recognised as the most accurate method [20–24], why not use the same strategy? After all, the ultimate goal is to superimpose two cephalograms at two time points, and tracing the important structure is just simplification of a cephalogram, which facilitates the operation of orthodontists because complex overlapping craniofacial structures, especially bilateral ones, are difficult for human eyes to distinguish.
Recent studies [25, 26] have proposed improved computer algorithms for feature matching, whose mission is to detect and match keypoints of the same or similar regions in multiple images taken at different viewpoints, under different illuminations, or at different magnifications. In comparison with the traditional manual process of superimposing the structure on two cephalograms, the computer algorithm bears many similarities and could be applicable as an automated superimposition method. One of the algorithms, Oriented FAST and Rotated BRIEF(ORB)[25], was shown to be time saving for the matching process, rotational invariant and noise immune. However, two aspects of this method should be improved for it to be practical for clinical applications: (1) the area for detecting and matching keypoints should be limited in the stable regions on the cephalograms; (2) to achieve accurate matching results, the matches should be not only abundant but also of high quality.
Currently, there is no study that describes a method for automated cephalometric structural superimposition. Therefore, the present study aimed to (1) establish a computer-aided automated method of structural superimposition for the anterior cranial base, maxilla and mandible and (2) evaluate the accuracy of automated structural superimposition based on free-hand tracing and superimposition.