DenseSeg: joint learning for semantic segmentation and landmark detection using dense image-to-shape representation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Semantic segmentation and landmark detection are crucial for medical image analysis.
• Deep learning excels at segmentation but struggles with landmark detection, traditionally a strength of shape-based methods.

Purpose of the Study:

• To propose a novel dense image-to-shape representation for joint learning of landmarks and semantic segmentation.
• To enable intuitive extraction of arbitrary landmarks by representing anatomical correspondences.

Main Methods:

• A fully convolutional architecture is employed for joint learning of semantic segmentation and landmark detection.
• The method utilizes a dense image-to-shape representation.
• Performance is benchmarked against state-of-the-art methods for segmentation and landmark detection.

Main Results:

• The method achieves comparable results to landmark detection baselines on thorax X-rays.
• It substantially surpasses landmark detection baselines on the more complex pediatric wrist dataset.
• Evaluation was performed on thorax X-rays and pediatric wrist bone datasets.

Conclusions:

• Dense geometric shape representation is advantageous for challenging landmark detection tasks.
• The proposed method outperforms previous state-of-the-art heatmap regression techniques.
• The approach allows for adding new landmarks without retraining, enhancing flexibility.