Contrastive Embeddind Learning Method for Respiratory Sound Classification

Abstract

Respiratory sound classification refers to identifying adventitious sounds from given recordings automatically. Due to the difficulty of collection and the expensive manual annotation, there are only limited samples available, which impacts on learning better models. Meanwhile, a majority of these models do not explicitly encourage intra-class compactness and inter-class separability between the learned embeddings, leading to the difficulty of identifying several samples and a reduced generalization performance. To address the problems, we propose a contrastive embedding learning method, where the input is a contrastive tuple. And the composite input strategy provides more possible network inputs. By the comparison among the samples in the tuple, we can learn the slight differences among the similar samples, and the easilyconfused samples are more likely to be identified. In the embedding space, we explicitly promote the intra-class compactness and inter-class separability, thereby the generalization performance is improved. Our method is evaluated on ICBHI 2017, and the classification score is increased from 75.61% of a conventional cross-entropy network to 78.18%, outperforming the state-of-the-art methods.

Example-based Explanations with Adversarial Attacks for Respiratory Sound Analysis

“…Evaluation Metrics. We report the unweighted average recall (UAR) as the generic classification benchmark instead of accuracy, as UAR can provide fairer evaluation of the models over the four classes than accuracy [9,22] in case of imbalance. It is also common to distinguish abnormal audio samples (i. e., crackles, wheezes, and both) from normal cases.…”

“…As a result, it is difficult for healthcare practitioners to fully trust the predictions if no explanation is available, especially when many respiratory sound classification results still have modest performance (e. g., the average score of around 50.16 % on the ICBHI 2017 dataset [8,2]). Existing works tried to mitigate this problem with data augmentation [9] to feed more data to DNNs. Nevertheless, the interpretability of a model is crucial in high-stake domains such as healthcare [10,11,12].…”

Example-based Explanations with Adversarial Attacks for Respiratory Sound Analysis

“…Evaluation Metrics. We report the unweighted average recall (UAR) as the generic classification benchmark instead of accuracy, as UAR can provide fairer evaluation of the models over the four classes than accuracy [8,20] in case of imbalance. It is also common to distinguish abnormal audio samples (i. e., crackles, wheezes, and both) from normal cases.…”

“…As a result, it is difficult for healthcare practitioners to fully trust the predictions if no explanation is available, especially when many respiratory sound classification results still have modest performance (e. g., the average score of around 50.16 % on the ICBHI 2017 dataset [7,2]). Existing works tried to mitigate this problem with data augmentation [8] to feed more data to DNNs. Nevertheless, the interpretability of a model is crucial in high-stake domains such as healthcare, where mistakes can cause significant damage [9,10].…”

SPRSound: Open-Source SJTU Paediatric Respiratory Sound Database

“…Feature Extraction Methods plays a key role in the sound analysis frameworks, reducing redundant information and improving the classification accuracy [59]. Various feature extraction methods have been explored in the state-of-the-art works, such as mel-frequency cepstral coefficient (MFCC) [8], [15], [28], [30], [60]- [65], log-mel spectrogram (Log-Mel) [59], [66], [67], mel spectrogram [34], [68]- [70] (Mel), and short-time fourier transform (STFT) spectrogram [31],…”