Kim June-Woo ; Yoon Chihyeon ; Toikkanen Miika ; Bae Sangmin ; Jung Ho-Young

Abstract

Deep generative models have emerged as a promising approach in the medicalimage domain to address data scarcity. However, their use for sequential datalike respiratory sounds is less explored. In this work, we propose astraightforward approach to augment imbalanced respiratory sound data using anaudio diffusion model as a conditional neural vocoder. We also demonstrate asimple yet effective adversarial fine-tuning method to align features betweenthe synthetic and real respiratory sound samples to improve respiratory soundclassification performance. Our experimental results on the ICBHI datasetdemonstrate that the proposed adversarial fine-tuning is effective, while onlyusing the conventional augmentation method shows performance degradation.Moreover, our method outperforms the baseline by 2.24% on the ICBHI Score andimproves the accuracy of the minority classes up to 26.58%. For thesupplementary material, we provide the code athttps://github.com/kaen2891/adversarial_fine-tuning_using_generated_respiratory_sound.