Improving Pediatric Pneumonia Diagnosis with Adult Chest X-ray Images Utilizing Contrastive Learning and Embedding Similarity

TL;DR

This work tackles pediatric pneumonia diagnosis by leveraging large-scale adult CXR data through a three-branch parallel framework that uses multi-positive contrastive learning and embedding similarity to reduce domain shift. By employing three ResNet-50–based paths (pediatric-only, adult-only, and joint), the method aligns classwise embeddings across domains, yielding an AUROC of $0.8464$ on pediatric tests—improving over the $0.8348$ achieved by naive joint training. The approach integrates projection heads, focal/classification losses, and embedding losses to cluster disease representations while diminishing domain bias, demonstrating potential for generalized CAD across pediatric and adult chest radiographs. These findings suggest practical pathways to scalable, cross-age pneumonia detection systems using existing adult datasets to augment pediatric performance.

Abstract

Despite the advancement of deep learning-based computer-aided diagnosis (CAD) methods for pneumonia from adult chest x-ray (CXR) images, the performance of CAD methods applied to pediatric images remains suboptimal, mainly due to the lack of large-scale annotated pediatric imaging datasets. Establishing a proper framework to leverage existing adult large-scale CXR datasets can thus enhance pediatric pneumonia detection performance. In this paper, we propose a three-branch parallel path learning-based framework that utilizes both adult and pediatric datasets to improve the performance of deep learning models on pediatric test datasets. The paths are trained with pediatric only, adult only, and both types of CXRs, respectively. Our proposed framework utilizes the multi-positive contrastive loss to cluster the classwise embeddings and the embedding similarity loss among these three parallel paths to make the classwise embeddings as close as possible to reduce the effect of domain shift. Experimental evaluations on open-access adult and pediatric CXR datasets show that the proposed method achieves a superior AUROC score of 0.8464 compared to 0.8348 obtained using the conventional approach of join training on both datasets. The proposed approach thus paves the way for generalized CAD models that are effective for both adult and pediatric age groups.

Figures (1)

• Figure 1: Overview of the proposed framework. The backbone models are based on ResNet-50 architecture. The adult and pediatric backbone models take adult and pediatric CXR images as input, respectively, while the common backbone model takes both adult and pediatric CXR images. Projection heads are used with the pooled global feature map to generate the embeddings. Three separate classifiers are utilized for predicting pathology probability and classification losses. Finally, the contrastive and embedding losses are utilized in embedding feature vector space to cluster the classwise embeddings, both intra- and inter-models, to reduce the impact of the domain gap.