TY - GEN
T1 - DiENTeS
T2 - 2nd International ToothFairy2: Multi-Structure Segmentation in CBCT Volumes, ToothFairy 2024, 3D Teeth Landmarks Detection Challenge, 3DTeethLand 2024, Semi-supervised Teeth Segmentation, STS 2024 held in conjunction with the 27th International Conference on Medical Image Computing and Computer-Assisted Intervention, MICCAI 2024
AU - Daza, Laura
AU - Schnabel, Julia
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.
PY - 2025
Y1 - 2025
N2 - Semantic segmentation is crucial for accurately identifying anatomical structures and pathological anomalies in medical images, playing a vital role in diagnostics, treatment planning, and disease progression monitoring. Despite significant advancements, the development of flexible and generalizable algorithms that can adapt to the diverse shapes, sizes, and textures of various anatomical regions remains challenging. In this work, we introduce the Dynamic ENTity Segmentation (DiENTeS) model, which leverages Local-global Transformers for 3D medical segmentation. Our model utilizes a transformer-based backbone to extract localized features and propagate them to form a comprehensive global representation. Additionally, we incorporate language features to guide the segmentation process, enabling the generation of specialized convolutional kernels for each category. This approach allows DiENTeS to tackle semantic segmentation as a class-agnostic entity segmentation problem. We validate our method using the ToothFairy2 Challenge, demonstrating its effectiveness in segmenting multiple structures in the maxillofacial region. We will make our code and models publicly available.
AB - Semantic segmentation is crucial for accurately identifying anatomical structures and pathological anomalies in medical images, playing a vital role in diagnostics, treatment planning, and disease progression monitoring. Despite significant advancements, the development of flexible and generalizable algorithms that can adapt to the diverse shapes, sizes, and textures of various anatomical regions remains challenging. In this work, we introduce the Dynamic ENTity Segmentation (DiENTeS) model, which leverages Local-global Transformers for 3D medical segmentation. Our model utilizes a transformer-based backbone to extract localized features and propagate them to form a comprehensive global representation. Additionally, we incorporate language features to guide the segmentation process, enabling the generation of specialized convolutional kernels for each category. This approach allows DiENTeS to tackle semantic segmentation as a class-agnostic entity segmentation problem. We validate our method using the ToothFairy2 Challenge, demonstrating its effectiveness in segmenting multiple structures in the maxillofacial region. We will make our code and models publicly available.
KW - Semantic segmentation
KW - ToothFairy2 Challenge
KW - multimodal segmentation
KW - vision transformers
UR - https://www.scopus.com/pages/publications/105006929893
U2 - 10.1007/978-3-031-88977-6_3
DO - 10.1007/978-3-031-88977-6_3
M3 - Conference contribution
AN - SCOPUS:105006929893
SN - 9783031889769
T3 - Lecture Notes in Computer Science
SP - 21
EP - 29
BT - Supervised and Semi-supervised Multi-structure Segmentation and Landmark Detection in Dental Data - MICCAI 2024 Challenges
A2 - Wang, Yaqi
A2 - Qian, Dahong
A2 - Wang, Shuai
A2 - Ben-Hamadou, Achraf
A2 - Pujades, Sergi
A2 - Lumetti, Luca
A2 - Bolelli, Federico
A2 - Grana, Costantino
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 6 October 2024 through 6 October 2024
ER -