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

BACKGROUND AND OBJECTIVE

Convolutional neural networks are widely used to detect radiological findings in chest radiographs. Standard architectures are optimized for images of relatively small size (for example, 224 × 224 pixels), which suffices for most application domains. However, in medical imaging, larger inputs are often necessary to analyze disease patterns. A single scan can display multiple types of radiological findings varying greatly in size, and most models do not explicitly account for this. For a given network, whose layers have fixed-size receptive fields, smaller input images result in coarser features, which better characterize larger objects in an image. In contrast, larger inputs result in finer grained features, beneficial for the analysis of smaller objects. By compromising to a single resolution, existing frameworks fail to acknowledge that the ideal input size will not necessarily be the same for classifying every pathology of a scan. The goal of our work is to address this shortcoming by proposing a lightweight framework for multi-scale classification of chest radiographs, where finer and coarser features are combined in a parameter-efficient fashion.

METHODS

We experiment on CheXpert, a large chest X-ray database. A lightweight multi-resolution (224 × 224, 448 × 448 and 896 × 896 pixels) network is developed based on a Densenet-121 model where batch normalization layers are replaced with the proposed size-specific batch normalization. Each input size undergoes batch normalization with dedicated scale and shift parameters, while the remaining parameters are shared across sizes. Additional external validation of the proposed approach is performed on the VinDr-CXR data set.

RESULTS

The proposed approach (AUC 83.27±0.17, 7.1M parameters) outperforms standard single-scale models (AUC 81.76±0.18, 82.62±0.11 and 82.39±0.13 for input sizes 224 × 224, 448 × 448 and 896 × 896, respectively, 6.9M parameters). It also achieves a performance similar to an ensemble of one individual model per scale (AUC 83.27±0.11, 20.9M parameters), while relying on significantly fewer parameters. The model leverages features of different granularities, resulting in a more accurate classification of all findings, regardless of their size, highlighting the advantages of this approach.

CONCLUSIONS

Different chest X-ray findings are better classified at different scales. Our study shows that multi-scale features can be obtained with nearly no additional parameters, boosting performance.