(주) 뷰노

JRC 2025

Deep Learning-based Lesion Segmentation for Patients with Usual Interstitial Pneumonia in High-Resolution CT: A pilot study

  • Apr. 2025

Abstract

Assessment of lung involvement in computed tomography (CT) images of patients with usual interstitial pneumonia (UIP) is crucial for management. However, due to the non-focal nature of lesions throughout the lung, developing a deep learning-based segmentation model requires extensive annotation, which is both time-consuming and costly. Therefore, a pilot study is required to estimate the budget for dataset collection when developing a UIP lesion segmentation model. In this pilot study, we annotated 155 slices containing lesions from high-resolution CT (HRCT) scans of 27 patients with UIP and used them to train a model, which was then evaluated on 56 slices from HRCT scans of 3 independent patients. A 2D Attention U-Net architecture was used for model development. Despite the relatively small dataset, the model achieved an average slice-level Dice coefficient of 0.819 on the test set, including normal slices. Additionally, at the case level, the three cases showed Dice coefficients of 0.837, 0.727, and 0.484. The Dice coefficient was relatively low for cases with less lung involvement. One notable observation is that the annotations were intentionally performed in an over-segmented manner, rather than tightly fitting the lung parenchyma boundaries, to reduce the burden on radiologists. Consequently, the model predictions followed the over-segmented annotations. Thus, we found that precise annotation within the lung region is crucial. In conclusion, this study demonstrates the feasibility of developing a model for lesion segmentation in CT scans of UIP patients using a small dataset, showing that reasonable results can be achieved with limited data.

Link

-

Author

Doohyun Park, Jung-Hyun Kang, Jonghun Jeong, Kyung Min Moon

Tag in

#medical_image

#VUNO Med®-LungCT AI™