Lightweight multi-scale classification of chest radiographs via size-specific batch normalization.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Radiology
• Computer-Aided Diagnosis

Background:

• Convolutional neural networks (CNNs) are standard for chest radiograph analysis.
• Existing CNNs often use fixed input sizes (e.g., 224x224 pixels), which may not be optimal for detecting findings of varying sizes.
• Different pathologies may require analysis at different image resolutions for accurate classification.

Purpose of the Study:

• To develop a lightweight, multi-scale framework for chest radiograph classification.
• To address the limitation of single-resolution input sizes in current CNN architectures.
• To efficiently combine features from different scales for improved detection of radiological findings.

Main Methods:

• A multi-resolution (224x224, 448x448, 896x896 pixels) Densenet-121 based network was developed.
• Size-specific batch normalization was implemented, with dedicated scale/shift parameters for each resolution.
• The model was trained and validated on the CheXpert and VinDr-CXR datasets.

Main Results:

• The proposed multi-scale approach achieved an AUC of 83.27±0.17, outperforming single-scale models.
• It demonstrated comparable performance to an ensemble of single-scale models but with significantly fewer parameters (7.1M vs. 20.9M).
• The model effectively leveraged features of different granularities for accurate classification of all findings, irrespective of size.

Conclusions:

• Chest X-ray findings exhibit varying optimal classification scales.
• Multi-scale feature extraction can significantly boost classification performance.
• This can be achieved with a parameter-efficient framework, offering practical advantages in medical imaging AI.