Sequential Patch Analysis Framework for Lung Disease Classification

Deep neural network (DNN) models have shown promise in lung disease classification using spectrogram transformations of respiratory sounds. However, these models typically process entire images, which can be computationally expensive, and they require large datasets for effective training. With the largest available respiratory dataset, ICBHI, containing only 920 recordings, training a robust DNN becomes challenging. To address this issue, this paper introduces a simplified DNN framework that incorporates novel techniques such as patch splitting and transfer learning. Specifically, the proposed framework splits the spectrogram image into a sequence of fixed-size patches, effectively reducing computational demands. These patches are processed with the position embeddings by DNN encoders to capture acoustic characteristics, leveraging a pre-trained ImageNet model to make efficient use of the limited dataset. Evaluation experiments on the ICBHI dataset show that the proposed method improves the state-of-the-art 8-class classification accuracy by 1.91%.