Enhancing Efficiency in CNN-Based Respiratory Sound Analysis through Temporal Self-Attention and Frequency Band
None:
Automating lung sound classification necessitates capturing both local spectral patterns and long-range temporal dependencies, a challenge for traditional models. While Convolutional Neural Networks (CNNs) excel at local feature extraction, they struggle with modeling long-range dependencies. Transformer-based models mitigate this limitation but incur high computational costs, rendering them impractical for real-time use in resource-constrained settings. To address these challenges, this paper proposes a lightweight CNN-based Temporal Self-Attention (CNN-TSA) network integrated with a Frequency Band Selection (FBS) strategy. The CNN-TSA efficiently captures temporal dependencies, while FBS eliminates irrelevant frequency bands, enhancing noise resilience and generalization. Experiments on the SPRSound 2022 dataset show that the proposed approach reduces computational costs by 50% without compromising performance, achieving state-of-the-art results. This efficient framework facilitates reliable real-time lung sound analysis, making it a promising solution for healthcare applications in resource-limited environments.
