Cephalometric landmark detection via global and local encoders and patch-wise attentions

Abstract

Cephalometric landmark detection performs an important role in a diagnostic measurements for orthodontic treatment plans. As manual depiction of landmarks is a time-consuming and tedious task, the development of an automated detection algorithm for daily clinics is in high demand. In this study, we propose a single-passing convolutional neural network that performs an accurate landmark detection in a hierarchical fashion. The proposed network first extracts global contexts by regressing initial positions of all the landmarks. Subsequently, local features are extracted from landmark-centered patches, which are obtained through global regression. The encoded global and local features are concatenated and weighed through a novel patch-wise attention module to weigh the relative importance. The experimental results demonstrate that our proposed local patch-wise attention mechanism performs a significant role in accurate detection. The proposed method outperformed other state-of-the-art methods by improving the successful detection rate by approximately 1 ti 2%. The proposed method suggests that a structured attention, which is developed in a patch-wise fashion, significantly enhances the local feature encoders to further improve the final accuracy. (c) 2021 Elsevier B.V. All rights reserved.