The paper presents a novel framework called TeethSEG for efficient instance segmentation of teeth in 2D intraoral images. The key highlights are:
Creation of the IO150K dataset: The authors created the largest open-source 2D intraoral image dataset, comprising over 150,000 images with professional annotations by orthodontists. The dataset covers a wide range of dental malformations.
Multi-Scale Aggregation (MSA) Blocks: TeethSEG utilizes MSA blocks to effectively aggregate visual semantics into trainable class embeddings at different scales. The MSA blocks employ a unique multi-head cross-gating mechanism to emphasize valuable components while maintaining the divergence between token embeddings.
Anthropic Prior Knowledge (APK) Layer: The APK layer incorporates human prior knowledge about tooth anatomy and positioning into the segmentation process, making the framework more interpretable and robust, especially in cases of tooth loss or abnormalities.
Permutation-based Upscaler: The authors introduce a permutation-based upscaler to generate clear segmentation edges and maintain rich local information in the image patch embeddings, addressing the limitations of previous transformer-based decoders.
The experiments demonstrate that TeethSEG outperforms state-of-the-art general-purpose segmentation models on dental image segmentation, both in independent and identically distributed (i.i.d.) test sets as well as out-of-distribution (o.o.d.) and RGB test sets.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Bo Zou,Shaof... at arxiv.org 04-02-2024
https://arxiv.org/pdf/2404.01013.pdfDeeper Inquiries