Transfer Learning with Limited Data for Mice Lung Segmentation in Synchrotron X-Ray Tomography

Area of Science:

• Medical Imaging
• Computational Biology
• Biomedical Engineering

Background:

• Accurate segmentation of lung structures is crucial for understanding lung development and disease.
• Traditional segmentation methods require large annotated datasets, which are difficult and costly to obtain, especially for complex small-scale biological samples like mouse lungs.
• Annotating mouse lungs from synchrotron radiation-based X-ray tomographic microscopy (SRXTM) images is particularly challenging due to their small size, complexity, and subtle boundaries.

Purpose of the Study:

• To develop a novel approach for segmenting mouse lungs in SRXTM images using transfer learning with limited annotated data.
• To address the challenges of unclear tissue-air demarcation and the scarcity of annotated datasets in medical imaging.
• To enable more efficient and accurate analysis of lung development, structural changes, and disease.

Main Methods:

• Utilized a U-net model pretrained on lung section images.
• Applied transfer learning to fine-tune the model for segmenting entire mouse lungs in SRXTM images.
• Evaluated the model's performance using a small dataset of 5 labeled entire lung image/mask pairs for training and 2 pairs for testing.

Main Results:

• The U-net model achieved high accuracy in segmenting lung sections, with a Dice Similarity Coefficient (DSC) of 0.9069 and a mean Intersection over Union (mIOU) of 0.895.
• After transfer learning for entire lung segmentation, the model maintained competitive performance with DSC = 0.8915 and mIOU = 0.8895.
• The method demonstrated effectiveness even with a minimal number of annotated images (5 pairs).

Conclusions:

• The proposed transfer learning approach is effective for segmenting mouse lungs in SRXTM images.
• This method significantly reduces the requirement for annotated data compared to traditional techniques.
• The findings highlight the potential of transfer learning to overcome data limitations in medical image analysis.