Published
The 80th Korean Congress of Radiology (KCR 2024)
Authors
Pin Jui Huang1 , Wen Tai1 , Yongxiang Wang1 , Dongmyung Shin2 , Minkyung Lee2
Affiliations
1MarketechInternationalCorporation, Chinese Taipei
2Radisen Co. Ltd.,Korea, Republic of
mklee317@gmail.com
*to be published soon.
Abstract
To evaluate the diagnostic performance of an AI algorithm that integrates deep learning-based denoising and classification models in detecting pneumonia on low-dose, noisy pediatric chest radiographs.
Reducing radiation exposure is crucial for pediatric patients. However, low-dose radiographs often contain noises, which can affect diagnostic accuracy. To address this, we utilized a pediatric pneumonia chest X-ray (PPCX) dataset to synthesize low-dose, noisy radiographs for AI-based pneumonia detection. A denoising AI model (DnCNN) was trained in a supervised manner using 220 radiographs randomly selected from the PPCX dataset. The original radiographs were used as targets and simulated noisy radiographs as training input. Synthetic Gaussian noises (σ_G∈ {5, 10, 15, 20}) were added to mimic thermal and electrical noises in low-dose X-ray images. Moreover, we also considered adding synthetic Poisson-Gaussian noises (σ_G=10, σ_P ∈ {10, 20, 30}) to simulate the random nature of photon detectors. After developing the denoising AI model, we integrated a classification AI model (ResNet-50) for pneumonia detection. This model was trained with am official training split of the PPCX dataset (3735 for pneumonia; 1295 for non-pneumonia), excluding images used in training the denoising model. While training, we froze the denoising model and added the samedegree of synthetic noises used for denoising AI development to the training samples. An official evaluation split of PPCX (375 for pneumonia; 225 for non-pneumonia) was used. Sensitivity, specificity, and area under curve of ROC curve (AUC-ROC) were measured using the test radiographs with different levels of synthetic noises.