TY - GEN

T1 - Assignment Theory-Augmented Neural Network for Dental Arch Labeling

AU - Dascalu, Tudor

AU - Ibragimov, Bulat

N1 - Publisher Copyright: © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023.

PY - 2023

Y1 - 2023

N2 - Identifying and detecting a set of objects that conform to a structured pattern, but may also have misaligned, missing, or duplicated elements is a difficult task. Dental structures serve as a real-world example of such objects, with high variability in their shape, alignment, and number across different individuals. This study introduces an assignment theory-based approach for recognizing objects based on their positional inter-dependencies. We developed a distance-based anatomical model of teeth consisting of pair-wise displacement vectors and relative positional scores. The dental model was transformed into a cost function for a bipartite graph using a convolutional neural network (CNN). The graph connected candidate tooth labels to the correct tooth labels. We re-framed the problem of determining the optimal tooth labels for a set of candidate labels into the problem of assigning jobs to workers. This approach established a theoretical connection between our task and the field of assignment theory. To optimize the learning process for specific output requirements, we incorporated a loss term based on assignment theory into the objective function. We used the Hungarian method to assign greater importance to the costs returned on the optimal assignment path. The database used in this study consisted of 1200 dental meshes, which included separate upper and lower jaw meshes, collected from 600 patients. The testing set was generated by an indirect segmentation pipeline based on the 3D U-net architecture. To evaluate the ability of the proposed approach to handle anatomical anomalies, we introduced artificial tooth swaps, missing and double teeth. The identification accuracies of the candidate labels were 0.887 for the upper jaw and 0.888 for the lower jaw. The optimal labels predicted by our method improved the identification accuracies to 0.991 for the upper jaw and 0.992 for the lower jaw.

AB - Identifying and detecting a set of objects that conform to a structured pattern, but may also have misaligned, missing, or duplicated elements is a difficult task. Dental structures serve as a real-world example of such objects, with high variability in their shape, alignment, and number across different individuals. This study introduces an assignment theory-based approach for recognizing objects based on their positional inter-dependencies. We developed a distance-based anatomical model of teeth consisting of pair-wise displacement vectors and relative positional scores. The dental model was transformed into a cost function for a bipartite graph using a convolutional neural network (CNN). The graph connected candidate tooth labels to the correct tooth labels. We re-framed the problem of determining the optimal tooth labels for a set of candidate labels into the problem of assigning jobs to workers. This approach established a theoretical connection between our task and the field of assignment theory. To optimize the learning process for specific output requirements, we incorporated a loss term based on assignment theory into the objective function. We used the Hungarian method to assign greater importance to the costs returned on the optimal assignment path. The database used in this study consisted of 1200 dental meshes, which included separate upper and lower jaw meshes, collected from 600 patients. The testing set was generated by an indirect segmentation pipeline based on the 3D U-net architecture. To evaluate the ability of the proposed approach to handle anatomical anomalies, we introduced artificial tooth swaps, missing and double teeth. The identification accuracies of the candidate labels were 0.887 for the upper jaw and 0.888 for the lower jaw. The optimal labels predicted by our method improved the identification accuracies to 0.991 for the upper jaw and 0.992 for the lower jaw.

KW - Assignment theory

KW - Dental instance classification

KW - Multi-object recognition

UR - http://www.scopus.com/inward/record.url?scp=85174721386&partnerID=8YFLogxK

U2 - 10.1007/978-3-031-43898-1_29

DO - 10.1007/978-3-031-43898-1_29

M3 - Article in proceedings

AN - SCOPUS:85174721386

SN - 9783031438974

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 295

EP - 304

BT - Medical Image Computing and Computer Assisted Intervention – MICCAI 2023 - 26th International Conference, Proceedings

A2 - Greenspan, Hayit

A2 - Greenspan, Hayit

A2 - Madabhushi, Anant

A2 - Mousavi, Parvin

A2 - Salcudean, Septimiu

A2 - Duncan, James

A2 - Syeda-Mahmood, Tanveer

A2 - Taylor, Russell

PB - Springer

T2 - 26th International Conference on Medical Image Computing and Computer-Assisted Intervention, MICCAI 2023

Y2 - 8 October 2023 through 12 October 2023

ER -