Distribution-guided Generative Replay with Semantic Prompts for Class-Incremental Chest X-ray Diagnosis


Jayant Mahawar (Indian Institue of Technology Jodhpur), Devi Prasad Maharathy (Indian Institute of Technology Jodhpur), Angshuman Paul (Indian Institute of Technology Jodhpur)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)Supplementary

Abstract

Existing class-incremental learning (CIL) methods perform poorly for the diagnosis of medical images. Some CIL approaches require storing exemplars, which raises privacy and storage concerns. Others rely on unconditioned generative replay, compromising discriminative power. To overcome these issues, we propose a new CIL framework for chest x-ray (CXR) diagnosis that combines prompt tuning with distribution-guided generative replay at the feature level. The prompts are initialized with semantically rich embeddings and refined through training. A variational autoencoder captures the feature distribution in latent space, enabling past knowledge retention without storing raw data. To balance stability and plasticity, parts of the network are frozen after the initial phase while others adapt to new classes. In each session, the model learns new classes using real data and replays the synthetic features of old ones to reduce forgetting. A classification module picks the closest class prompt using cosine similarity. We evaluate our method on public CXR datasets. It outperforms prior CIL methods in accuracy and retention. We achieve up to 9% improvement in average accuracy compared to SOTA methods.

Citation

@inproceedings{Mahawar_2025_BMVC,
author    = {Jayant Mahawar and Devi Prasad Maharathy and Angshuman Paul},
title     = {Distribution-guided Generative Replay with Semantic Prompts for Class-Incremental Chest X-ray Diagnosis},
booktitle = {36th British Machine Vision Conference 2025, {BMVC} 2025, Sheffield, UK, November 24-27, 2025},
publisher = {BMVA},
year      = {2025},
url       = {https://bmva-archive.org.uk/bmvc/2025/assets/papers/Paper_122/paper.pdf}
}


Copyright © 2025 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection