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

Due to the small scale and complexity of the acinus, the demarcation between tissue and air is often unclear, making segmentation challenging. Accurate segmentation is essential for analyzing lung development, structural changes, and functional roles, offering insights into lung growth and disease. Traditional methods require large, annotated datasets, which are difficult to obtain, especially in medical imaging, where annotations are time-consuming and costly. Annotating mice lungs imaged by synchrotron radiation-based X-ray tomographic microscopy (SRXTM) is particularly difficult due to their small size, complexity, and subtle boundaries.

This paper presents a novel approach using transfer learning with limited data for segmenting mice lungs in SRXTM images. A U-net model, evaluated on a dataset of 59 image/mask pairs of lung sections, was pretrained on lung section images and then used to segment entire lungs in images. Segmentation accuracy improved through fine-tuning with transfer learning, using only 5 labeled entire lung image/mask pairs, with 2 pairs for testing. This highlights the potential of transfer learning in addressing the challenge of limited annotated data.

The results show that U-net applied on the lung sections images achieves accurate prediction of the lung section masks, with the following metrics DSC = 0.9069, mIOU = 0.895, and BCE = 0.141. After transfer learning to entire lung images, segmentation results were DSC = 0.8915, mIOU = 0.8895, and BCE = 0.141.

Experimental results demonstrate the effectiveness of the proposed method, achieving competitive segmentation performance compared to traditional approaches, while requiring significantly fewer annotated images.