Nonparametric Bayesian transfer learning for robust cardiopulmonary diseases classification in X-ray images

Deep learning has revolutionized the detection of cardiopulmonary diseases by using readily available X-ray images. Transfer learning offers an exciting avenue for accelerating progress in this field, particularly when large training datasets are scarce. However, difficulties arise when transferring...

Full description

Saved in:

Bibliographic Details
Published in:	Knowledge-based systems Vol. 326; p. 114034
Main Authors:	Haftu, Kibrom, Assabie, Yaregal
Format:	Journal Article
Language:	English
Published:	Elsevier B.V 27.09.2025
Subjects:	Cardiopulmonary disease classification Deep learning Deep transfer learning Indian buffet process Nonparametric Bayesian priors Variational autoencoder Deep learning Variational autoencoder Indian buffet process Cardiopulmonary disease classification Nonparametric Bayesian priors Deep transfer learning
ISSN:	0950-7051
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	Deep learning has revolutionized the detection of cardiopulmonary diseases by using readily available X-ray images. Transfer learning offers an exciting avenue for accelerating progress in this field, particularly when large training datasets are scarce. However, difficulties arise when transferring knowledge from one domain to another unrelated task, potentially harming model performance. Therefore, we propose a novel nonparametric Bayesian statistical model to investigate the effectiveness of transfer learning on radiographic images. The proposed model comprises of two main components: deep transfer learning and classification. The deep transfer learning component extracts domain-invariant discriminating features using an Indian buffet process-driven variational autoencoder. This Bayesian nonparametric model enables flexible modeling of networks with potentially unbounded sizes while simultaneously capturing complex structural patterns and regularities within the data. The classification component further fine-tunes these features using a supervised algorithm rather than the current approaches that use a single feature space represented by the last fully connected layer of the convolutional neural networks across all conditions. Our model achieved a mean area under the curve (AUC) score of 88.01% for 14 cardiopulmonary diseases in the NIH chest radiograph dataset, outperforming the existing state-of-the-art methods. Validation of the collected external data demonstrates the generalizability of the model.
ISSN:	0950-7051
DOI:	10.1016/j.knosys.2025.114034