Table 3 Deep learning methods in lung sound analysis
From: Deep learning-based lung sound analysis for intelligent stethoscope
Task | Year | Ref | Basic method | Dataset | Outcome |
|---|---|---|---|---|---|
ASD | 2013 | [126] | Wavelets, FNN | Self-collected, 13 healthy, 13 pathological | Normal or crackle: ACC—71.55% |
2014 | [117] | Wavelets, FNN | Lehrer's dataset | Normal, wheeze or crackle: ACC—99.26% | |
2016 | [132] | MFCCs, LFCCs, FNN | RALE database, IIT Kharagpur, Salt Lake, Kolkata | Normal, wheeze or crackle: ACC—97.61%, SEN—97.41%, SPE—98.33% | |
2018 | [33] | Spectrogram, CNN | RALE dataset | Coarse crackle, fine crackle, polyphonic wheeze, monophonic wheeze, normal, squawk, stridor: ACC—95.56% | |
2018 | [127] | MFCCs, RNN | Self-collected, 10 healthy, 5 idiopathic pulmonary fibrosis | Inspiration: F1—87%; Expiration: F1—84.6%; Crackles: F1—72.1% | |
2018 | [147] | LSTM | ICBHI 2017 | Normal, crackles, wheezes, and both crackles and wheezes: SEN—58.4%, SPE—73.0%, AS—65.7% | |
2019 | [133] | Spectrogram, CNN | Self-collected, 50 pediatric patients | Wheeze, rhonchi, fine crackle, or coarse crackles: recall—76.5%, precision—53.0%, SPE—83.6%, F1—62.5% | |
2019 | [148] | Spectrogram, CNN | CBHI 2017 | Normal, crackles, wheezes, and both crackles and wheezes: SEN—31.12%, SPE—68.20%, AS—50.16% | |
2020 | [23] | Mel spectrogram, CNN | ICBHI 2017 | Normal, crackles, wheezes, and both crackles and wheezes: SEN—48.63%, SPE—84.14%, AS—66.38% | |
2020 | [131] | Mel spectrogram, CNN | ICBHI 2017 | Normal, crackles, wheezes, and both crackles and wheezes: AS—78.4%; normal and abnormal: AS—83.7% | |
2020 | [136] | Mel Spectrogram, CNN | ICBHI 2017 | Normal, wheeze or crackle: ACC—98.6%, F1—98.4% | |
2020 | [149] | Spectrogram, LSTM, CNN, autoencoder | Self-collected, 22 patients | Inspiration or expiration: ACC—92% | |
2020 | [150] | Spectrogram, CNN | Self-collected, 25 pediatric patients | Normal, wheeze or crackle: crackle PPA—95%, NPA—99%; wheeze PPA—90%, NPA—97% | |
2020 | [151] | Spectrogram, CNN | RALE database, Think labs Lung sound library | Normal, crackles, wheezes, or rhonchi: ACC—83.78% | |
2021 | [22] | Spectrogram, CNN | ICBHI 2017 | Normal, crackles, wheezes, and both crackles and wheezes: SPE—85.44%, SEN—70.93%, AS—78.18% | |
2021 | [72] | Mel spectrogram, CNN | ICBHI 2017 | Normal, crackles, wheezes, and both crackles and wheezes: SPE—72.3%, SEN—40.1%, AS—56.2%; Normal or abnormal: SPE—80.9%, SEN—73.1%, AS—77.0% | |
2021 | [77] | CNN | Self-collected, 1918 respiratory sound record | Normal, abnormal (crackles, wheezes, rhonchi): ACC—84.8%, precision—81.4%, recall—81.7%, F1—81.4% | |
2021 | [152] | Spectrogram, CNN, autoencoder | ICBHI 2017 | Normal, crackles, wheezes, and both crackles and wheezes: SPE—69%, SEN—29%, AS—49% | |
2023 | [82] | Spectrogram, CNN | Self-collected, 105 health, 189 patients | Normal, crackles, or rhonchi: ACC—83% | |
2022 | [130] | Mel spectrogram, CNN | ICBHI 2017 | Normal, crackles, wheezes, and both crackles and wheezes: ACC—84.7%, SEN—84.5%, SPE—84.9%, precision—84.4%, recall—89.0%, F1—86.6% | |
2022 | [134] | Spectrogram, CNN | ICBHI 2017 | Normal, crackles, wheezes, and both crackles and wheezes: SPE—85.6%, SEN—29%, AS—57.3% | |
2022 | [135] | Mel spectrogram, TCN | ICBHI 2017 | Normal, crackles, wheezes, and both crackles and wheezes: SPE—86.1%, SEN—65.3%, AS—75.7% | |
2022 | [137] | LSTM, CNN | ICBHI 2017 | Normal, crackles, wheezes, and both crackles and wheezes: SPE—82.46%, SEN—47.37%, AS—64.92% | |
2022 | [146] | Spectrogram, Mel spectrogram, CNN | ICBHI 2017 | Crackles or others: ACC—86.4%; Wheezes or others: ACC—78.2%; Crackles, wheezes, or others: ACC—84.5% | |
RDR | 2013 | [153] | Statistical feature, FNN | Self-collected, 27 healthy and 33 tuberculosis | Healthy or pulmonary tuberculosis subject: ACC—73% |
2014 | [138] | FNN | Self-collected, 10 healthy and 20 pathological | Normal and abnormal subject: ACC—92.86%, SEN—86.30%, SPE—86.90% | |
2018 | [27] | Deep belief networks | RespiratoryDatabase@TR | Risk level or interior level: ACC—95.84%, SEN—93.34%, SPE—93.65% | |
2018 | [91] | Extreme learning machines | RespiratoryDatabase@TR | COPD or health: ACC—92.22%, SEN—89.44%, SPE—95.00% | |
2018 | [154] | Deep extreme learning | RespiratoryDatabase@TR | COPD or health: ACC—95.0%, SEN—93.33%, SPE—93.53% | |
2019 | [123] | Spectrogram, CNN | ICBHI 2017 | All diseases classification: ACC—97% | |
2020 | [155] | Boltzmann machines | RespiratoryDatabase@TR | COPD or healthy subject: ACC -93.67%, SEN—91%, SPE—96.33% | |
2020 | [122] | Convolutional RNN | Self-collected, 16 healthy and 7 pulmonary fibrosis | Health or idiopathic pulmonary fibrosis: precision—100%, SEN—85.9%, F1—92.4% | |
2020 | [139] | Mel spectrogram, CNN | ICBHI 2017 | Non-COPD, COPD, or healthy subject: SEN—98.5%, SPE—99.0%, AS—98.7% | |
2020 | [143] | Extreme learning machines | RespiratoryDatabase@TR | Five severity degrees of COPD: ACC—94.31%, SEN—94.28%, SPE—98.76% | |
2020 | [156] | Statistical feature, FNN | ICBHI 2017 | Health or diseases: ACC—82%, precision—87% | |
2021 | [140] | EMD, wavelet, CNN | ICBHI 2017 | Non-COPD, COPD, or healthy subject: precision—98.90%, recall—98.90%, ACC—98.92%, F1—98.90%; Six diseases: precision—98.70%, recall—98.27%, ACC—98.70%, F1—98.47% | |
2021 | [144] | Deep belief network | RespiratoryDatabase@TR | Mild, moderate, or severe COPD: ACC—71.74%, SEN—70.08%, SPE—73.53% | |
2022 | [125] | CNN, LSTM | Self-collected, 103 patients | Health or five diseases: ACC—98.16%, SEN—90.06%, SPE—98.61%, precision—92.13% | |
2022 | [141] | Wavelet, CNN, LSTM | ICBHI 2017 | Health, COPD, asthma, or pneumonia: ACC—88.86%; Health, COPD, or non-COPD: ACC—66.54% | |
2022 | [142] | FNN, CNN, LSTM | ICBHI 2017 | URTI, COPD, pneumonia, and bronchiolitis and healthy: ACC—94%, precision—94%, recall—94%, F1—93% | |
2022 | [157] | Statistical feature, FNN | ICBHI 2017 | Health or diseases: SPE—97.6%, SEN—98.2% | |
2023 | [10] | CNN, LSTM | Self-collected, 198 patients | Disease, symptom relief, or health: 1) subject-dependent: SEN—96.98%, SPE—97.43%, AS—97.20% 2) subject-independent: SEN—43.26%, SPE—39.61%, AS—41.44% | |
2023 | [92] | Mel spectrogram, pretrained MobileNet-V1 | RespiratoryDatabase@TR | Risk level or interior level: ACC—99.25%, SEN—99.18%, SPE—99.36%; Five severity degrees of COPD: ACC—96.14%, SEN—95.94%, SPE—98.89% | |
2023 | [129] | CNN | Self-collected, 126 subject | Health, asthma, COPD, ILD, pneumonia, bronchiectasis: precision—92.81%, SEN—92.22%, SPE—98.50% | |
ASD, RDR | 2019 | [26] | MFCCs, LSTM | ICBHI 2017 | Normal, crackles, wheezes, and both crackles and wheezes: ACC—74%, SPE—85%, SEN—62%, AS—74%; Normal or abnormal: ACC—81%; Health or diseases: ACC—99%, SPE—82%, SEN—99%, AS—91%; Health, COPD, or non-COPD: ACC—98%, SPE—82%, SEN—98%, AS—90% |
2020 | [25] | Spectrogram, CNN | ICBHI 2017 | Normal, crackles, wheezes, and both crackles and wheezes: SPE—89%, SEN—72%, AS—80%; Normal or abnormal: SPE—89%, SEN—82%, AS—85%; Health or diseases: SPE—71%, SEN—99%, AS—85%; Health, COPD, or non-COPD: SPE—71%, SEN—95%, AS—83% | |
2021 | [124] | Spectrogram, CNN | ICBHI 2017 | Normal, crackles, wheezes, and both crackles and wheezes: SPE—90%, SEN—68%, AS—79%; Normal or abnormal: SPE—90%, SEN—78%, AS—84%; Health or diseases: SPE—86%, SEN—98%, AS—92%; Health, COPD, or non-COPD: SPE—86%, SEN—95%, AS—91% | |
2021 | [128] | Mel spectrograms, CNN | ICBHI 2017 | Normal, crackles, or wheezes: SPE—82%, SEN—61%, AS—72%; COPD, healthy, and pneumonia: SPE—92%, SEN—98%, AS—95% | |
2021 | [145] | Spectrogram, Inception | ICBHI 2017 | Normal, crackles, wheezes, and both crackles and wheezes: SPE—73%, SEN—30%, AS—52%; Health, COPD, or non-COPD: SPE—100%, SEN—75%, AS—85% | |
2022 | [24] | Spectrogram, CNN | ICBHI 2017 | Normal, crackles, wheezes, and both crackles and wheezes: SPE—78.55%, SEN—35.97%, AS—35.97%; Normal or abnormal: SPE—79.34%, SEN—50.14%, AS—64.74%; Health, COPD, or non-COPD: SPE—91.77%, SEN—93.68%, AS—92.72%; Health or diseases: SPE—91.77%, SEN—95.76%, AS—93.77% |