Attention-Guided Multi-Scale CNN Network for Cervical Vertebral Maturation Assessment from Lateral Cephalometric Radiography

International Workshop on Bayesian Inference and Maximum Entropy Methods in Science and Engineering (MaxEnt), pp. 26

Abstract

Accurate determination of skeletal maturation indicators is crucial in the orthodontic process. Chronologic age is not a reliable skeletal maturation indicator, thus physicians use bone age. In orthodontics, the treatment timing depends on Cervical Vertebral Maturation (CVM) assessment. Determination of CVM degree remains challenging due to the limited annotated dataset, the existence of significant irrelevant areas in the image, the huge intra-class variances, and the high degree of inter-class similarities. To address this problem, researchers have started looking for external information beyond current available medical datasets. This work utilizes the domain knowledge from radiologists to train neural network models that can be utilized as a decision support system. We proposed a novel supervised learning method with a multi-scale attention mechanism, and we incorporated the general diagnostic patterns of medical doctors to classify lateral X-ray images as six CVM classes. The proposed network highlights the important regions, surpasses the irrelevant part of the image, and efficiently models long-range dependencies. Employing the attention mechanism improves both the performance and interpretability. In this work, we used additive spatial and channel attention modules. Our proposed network consists of three branches. The first branch extracts local features, and creates attention maps and related masks, the second branch uses the masks to extract discriminative features for classification, and the third branch fuses local and global features. The result shows that the proposed method can represent more discriminative features, therefore, the accuracy of image classification is greater in comparison to in backbone and some attention-based state-of-the-art networks.

Authors 5

1. Hamideh Manoochehri

   Amirkabir University of Technology

   Affiliation as printed

   Department of Electrical Engineering, Amirkabir University of Technology, Tehran 1591634311, Iran

2. Seyed Ahmad Motamedi corresponding

   Amirkabir University of Technology

   Affiliation as printed

   Department of Electrical Engineering, Amirkabir University of Technology, Tehran 1591634311, Iran

3. Ali Mohammad‐Djafari corresponding
   Affiliation as printed

   International Science Consulting and Training (ISCT), 91440 Bures-sur-Yvette, France

4. Masrour Makaremi

   Université de Bordeaux

   Affiliation as printed

   Departement Dentofacial Orthopedics, UFR des Sciences Odontologiques, 146, Rue Léo-Saignat, CEDEX, 33076 Bordeaux, France

5. Alireza Vafaei Sadr Aachen Institute of Pathology

   Universitätsklinikum Aachen · RWTH Aachen University

   Affiliation as printed

   Institute of Pathology, RWTH Aachen University Hospital, 52074 Aachen, Germany