Baseline Method - DENTEX - MICCAI23

The DENTEX challenge has concluded. We deeply appreciate your participation. Results will be showcased at MICCAI 2023 in Vancouver, Canada on October 8. Please check the Final Announcements below for more information.

DENTEX CHALLENGE 2023

Dental Enumeration and Diagnosis on Panoramic X-rays¶

BASELINE METHOD¶

As our baseline approach, we present the HierarchicalDet framework for multi-label, hierarchically labeled tooth detection, as described in Hamamci et. al, 2023. This baseline method employs a diffusion-based object detection model, incorporating a denoising diffusion process to refine initial noisy bounding boxes into precise object boxes.

The baseline method is composed of an image encoder, responsible for extracting high-level features from input images, and a detection decoder that refines the initial noisy boxes using these features. We adopt the same diffusion process as in DiffusionDet, which formulates object detection as a denoising diffusion process transitioning from noisy boxes to object boxes.

To augment the baseline model's performance, we integrate hierarchical learning architecture that leverages hierarchically annotated data and a novel noisy box manipulation technique. This hierarchical learning architecture enables us to utilize previously inferred boxes for enhanced detection accuracy. The innovative noisy box manipulation technique facilitates efficient learning from partial annotations.

Furthermore, our baseline method incorporates multi-label object detection using a customized Detectron2 library to accommodate partial annotations. We selectively freeze the classification heads corresponding to unlabeled classes and exploit the full spectrum of accessible information to bolster our model's capacity to manage partially labeled data.

All codes related to our baseline method, HierarchicalDet, can be found on Github.

Fig. 3. Framework of the baseline method, HierarchicalDet, which relies on a hierarchical learning approach utilizing a combination of multi-label detection, bounding box manipulation, and weight transfer.