Multi-scale consistent self-training network for semi-supervised orbital tumor segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Orbital tumor diagnosis is crucial for eye health.
• Accurate segmentation of orbital tumors in CT images is challenging due to variations in size and shape.
• Limited annotated data further complicates automated segmentation tasks.

Purpose of the Study:

• To develop a robust semi-supervised segmentation method for orbital tumors in CT images.
• To address the challenges posed by diverse tumor characteristics and limited annotations.
• To improve the accuracy and generalization of orbital tumor segmentation models.

Main Methods:

• Proposed a multi-scale consistent self-training network (MSCINet) for semi-supervised orbital tumor segmentation.
• Utilized semantic-invariance features and enforced prediction consistency across different image scales.
• Incorporated an iterative self-training strategy with uncertainty filtering to refine pseudo-labels and minimize error accumulation.

Main Results:

• Developed two new datasets: Orbtum-B (binary segmentation) and Orbtum-M (multi-organ segmentation).
• The proposed MSCINet achieved state-of-the-art performance on both datasets.
• Evaluated on 55 patients with 602 2D images, demonstrating superior segmentation capabilities.

Conclusions:

• Developed a novel semi-supervised segmentation method specifically for orbital tumors.
• The proposed method exhibits excellent performance, outperforming previous semi-supervised algorithms.
• The approach is well-suited for the unique characteristics of orbital tumors, enhancing diagnostic potential.