Label efficient segmentation of single slice thigh CT with two-stage pseudo labels

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Accurate segmentation of muscle, bone, and fat in thigh images is crucial for body composition analysis.
• Deep learning models excel at medical image segmentation but typically require large annotated datasets.
• Manual annotation is costly and time-consuming, posing a challenge for training deep learning models with limited data.

Purpose of the Study:

• To develop and evaluate a two-stage deep learning pipeline for thigh and lower leg segmentation.
• To address the challenge of training deep learning models with limited human-labeled data for body composition quantification.
• To improve the efficiency and reduce the cost of creating annotated datasets for medical image segmentation.

Main Methods:

• A two-stage deep learning pipeline inspired by transfer learning was proposed.
• The first stage involved generating pseudo-labels for thigh images using handcrafted approaches based on CT intensity and morphology.
• These pseudo-labels trained deep neural networks from scratch, followed by fine-tuning with limited expert human labels.

Main Results:

• The framework achieved an average Dice Similarity Coefficient (DSC) of 0.927 for thigh segmentation across muscle, internal bone, cortical bone, subcutaneous fat, and intermuscular fat.
• Generalizability was tested on lower leg images, yielding an average DSC of 0.823.
• The method demonstrated effective segmentation performance on 73 thigh CT images.

Conclusions:

• Approximated handcrafted pseudo-labels can serve as effective initialization for deep neural networks.
• This approach significantly reduces the need for extensive human expert-labeled data.
• The proposed method maximizes the utility of available expert-labeled data for medical image segmentation tasks.