Cartilage Segmentation in High-Resolution 3D Micro-CT Images via Uncertainty-Guided Self-training with Very Sparse Annotation

Abstract

Craniofacial syndromes often involve skeletal defects of the head. Studying the development of the chondrocranium (the part of the endoskeleton that protects the brain and other sense organs) is crucial to understanding genotype-phenotype relationships and early detection of skeletal malformation. Our goal is to segment craniofacial cartilages in 3D micro-CT images of embryonic mice stained with phosphotungstic acid. However, due to high image resolution, complex object structures, and low contrast, delineating fine-grained structures in these images is very challenging, even manually. Specifically, only experts can differentiate cartilages, and it is unrealistic to manually label whole volumes for deep learning model training. We propose a new framework to progressively segment cartilages in high-resolution 3D micro-CT images using extremely sparse annotation (e.g., annotating only a few selected slices in a volume). Our model consists of a lightweight fully convolutional network (FCN) to accelerate the training speed and generate pseudo labels (PLs) for unlabeled slices. Meanwhile, we take into account the reliability of PLs using a bootstrap ensemble based uncertainty quantification method. Further, our framework gradually learns from the PLs with the guidance of the uncertainty estimation via self-training. Experiments show that our method achieves high segmentation accuracy compared to prior arts and obtains performance gains by iterative self-training.

Keywords: Cartilage segmentation; Sparse annotation; Uncertainty.

Fig. 1.

Examples of micro-CT images of stained mice. (a) A raw 3D image and its manual annotation. The shape variations are large: the front nasal cartilage is relatively small (i.e., 300²); the cranial vault is very big (i.e., 900 × 500) but extremely thin like a half-ellipsoid surface. (b) A 2D slice from the nasal cartilage (top) and its associated label (bottom); the image contrast is low and there are many hard mimics in surrounding areas. (c) Two 2D slices from the cranial vault (top) and their associated labels (bottom); the cartilage is very thin. Best viewed in color.

Fig. 2.

An overview of our proposed framework.

Fig. 3.

The network architecture of our proposed method, K-head FCN. The output layer branches out to K bootstrap heads and an extra log-variance output.

Fig. 4.

Qualitative examples: (a) Raw subregions; (b) ground truth; (c) U-Net* (TL); (d) K-head FCN (TL); (e) K-head FCN-R3-U (TL∪PL). (XX) = (trained using XX).

Fig. 5.

Visualization of uncertainty. From left to right: a raw image region, ground truth, prediction result, estimated epistemic uncertainty, and estimated aleatoric uncertainty. Brighter white color means higher uncertainty.