(주) 뷰노

KATRDIC 2025 (한국결핵및호흡기학회)

Improved Deep Learning-Based Computer-Aided Diagnosis for Usual Interstitial Pneumonia Classification in Chest CT scan

  • Oct. 2025
  • by Kyoung Min Moon et. al.

Background

Usual Interstitial Pneumonia (UIP) is a fibrosing interstitial lung disease characterized by specific radiologic features, such as honeycombing on high-resolution computed tomography (HRCT). Accurate diagnosis is critical for prognosis and treatment decisions but remains challenging. This study aimed to develop and validate an improved deep learning-based computer-aided diagnosis (CAD) system for differentiating UIP from non-UIP cases using HRCT.

 

Methods

A total of 503 axial HRCT scans were retrospectively collected from a tertiary hospital in Korea, including 381 UIP and 122 non-UIP cases. Lung segmentation was performed using the Segment Anything Model (SAM), guided by prompt-based masks. Preprocessing included lung region extraction, field segmentation, window-level adjustment, and multi-slice concatenation for spatial context. A modified ResNet50_2D architecture was trained to classify UIP versus non-UIP. Model performance was assessed on an independent validation set (40 UIP and 40 non-UIP) using area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, and accuracy.

 

Results

The CAD system achieved an AUROC of 0.77. At the optimal cutoff determined by the maximum Youden index, the model yielded a sensitivity of 0.999, specificity of 0.450, and overall accuracy of 0.725. All UIP cases were correctly identified (TP = 40, FN = 0), while among non-UIP cases, 18 were correctly classified (TN = 18) and 22 were misclassified as UIP (FP = 22).

 

Conclusions

The proposed CNN-based CAD system with advanced segmentation and preprocessing demonstrated promising performance in UIP classification on HRCT. By achieving near-perfect sensitivity, the system could support radiologists in reliably identifying UIP, though specificity requires further improvement. Ongoing external validation and clinical integration studies will determine its applicability in real-world practice.

Tag in

#medical_image