(주) 뷰노

KCR 2025

Improved Deep Learning-Based CAD System for Usual Interstitial Pneumonia Classification in Chest HRCT

  • Sep. 2025

PURPOSE :
Usual Interstitial Pneumonia (UIP) is a chronic fibrosing interstitial lung disease characterized by specific radiologic patterns, including honeycombing on high-resolution computed tomography (HRCT). Accurate identification of UIP is critical for prognosis and treatment planning but remains a diagnostic challenge. This study aimed to develop and enhance a deep learning-based computer-aided diagnosis (CAD) system to classify UIP and non-UIP cases in chest HRCT using a convolutional neural network (CNN) approach.

 

METERIALS AND METHOD :
A total of 503 axial HRCT scans were retrospectively collected from a Korean tertiary hospital, comprising 381 UIP and 122 non-UIP cases. For lung segmentation, we utilized the Segment Anything Model (SAM) developed by Meta AI (Kirillov, Alexander, et al., 2023), which was guided by prompt-based segmentation masks. The input images underwent a preprocessing pipeline including: (1) lung region extraction from HRCT using SAM, (2) segmentation of lung fields, (3) window-level adjustment focused on the lung parenchyma, and (4) multi-slice concatenation for spatial context enhancement. For classification, a modified ResNet50_2D architecture (He, Kaiming, et al., 2016.) was employed to distinguish between UIP and non-UIP cases using the preprocessed lung patch inputs. Model performance was evaluated on a separate validation set using AUROC, sensitivity, specificity, and overall accuracy as primary metrics.

 

RESULTS :
The classification model was evaluated on an independent validation set comprising 80 cases (40 UIP and 40 non-UIP). The model achieved an AUROC of 0.77, indicating good overall discriminatory power. At the optimal threshold determined by the maximum Youden index, the model achieved a sensitivity of 0.999, specificity of 0.450, and an overall accuracy of 0.725. The confusion matrix showed that all 40 UIP cases were correctly identified (TP = 40, FN = 0), while among the non-UIP cases, 18 were correctly classified (TN = 18) and 22 were falsely identified as UIP (FP = 22). 

 

CONCLUSION :
We present an improved CNN-based CAD system for UIP classification in chest HRCT. The enhanced segmentation and classification pipeline demonstrated promising diagnostic performance. This system has the potential to assist radiologists in early and accurate detection of UIP, thereby improving clinical decision-making. Further external validation and integration into clinical workflows are ongoing to assess its real-world applicability.
 

Tag in

#VUNO Med®-LungCT AI™