Deep-Learning Segmentation of Bleomycin-Induced Pulmonary Fibrosis in Rats Using U-Net 3 + by 3D UTE-MRI

Abstract

This study utilized the U-Net 3 + neural network to develop an algorithm for automatic lung segmentation in laboratory rats, identifying corresponding pathologies, particularly pulmonary fibrosis induced by intratracheal administration of bleomycin. MR images of rat lungs were obtained in 30 days after initialization of the fibrosis at 7 T using ultra-short echo time (UTE) pulse sequence. Initially, lung and pathology masks were highlighted manually, and then they were subsequently used to train the neural network. The proposed algorithm operates step by step, firstly segmenting the lungs and then detecting pathologies within them. The metric results demonstrate a good agreement between manual and automatic segmentation, with Dice Similarity Coefficient (DSC) = 0.93 ± 0.05 and Intersection over Union (IoU) = 0.83 ± 0.19 for the lungs, and DSC = 0.72 ± 0.19, IoU = 0.54 ± 0.22 for pulmonary fibrosis. The authors noted high accuracy in lung segmentation and the ability to effectively differentiate lung pathologies from surrounding normal tissues with minor inaccuracies in the shape and size of pathologies.