Source-linked AI summary
RespireNet: A Deep Neural Network for Accurately Detecting Abnormal Lung Sounds in Limited Data Setting
Siddhartha Gairola, Francis Tom, Nipun Kwatra, Mohit Jain
TL;DR
RespireNet addresses the challenge of training deep neural networks on the small ICBHI respiratory-sound dataset. It combines a simple CNN with data-efficient techniques and achieves new state-of-the-art results on the ICBHI 2-class and 4-class tasks, including a 2.2% gain over the prior state of the art on the 80-20 split.
Problem
The ICBHI dataset has only 6,898 breathing cycles, which is small for training data-hungry DNNs.
Source: §1 · p. 2, §2.1 · p. 2Method
Results
Takeaways & Limitations
The proposed techniques are orthogonal to network architecture and should be easy to incorporate into other frameworks.
Source: §1 · p. 2, §5 · p. 4The 4-class performance limit is attributed mainly to ICBHI's small size and device variation; results also vary across the nonstandard 80-20 splits.
Source: §5 · p. 4
Abstract
from arXiv · showhide
Auscultation of respiratory sounds is the primary tool for screening and diagnosing lung diseases. Automated analysis, coupled with digital stethoscopes, can play a crucial role in enabling tele-screening of fatal lung diseases. Deep neural networks (DNNs) have shown a lot of promise for such problems, and are an obvious choice. However, DNNs are extremely data hungry, and the largest respiratory dataset ICBHI has only 6898 breathing cycles, which is still small for training a satisfactory DNN model. In this work, RespireNet, we propose a simple CNN-based model, along with a suite of novel techniques -- device specific fine-tuning, concatenation-based augmentation, blank region clipping, and smart padding -- enabling us to efficiently use the small-sized dataset. We perform extensive evaluation on the ICBHI dataset, and improve upon the state-of-the-art results for 4-class classification by 2.2%
1. INTRODUCTION
Respiratory sound analysis could support screening and diagnosis, but manual interpretation requires trained professionals and can vary between listeners. RespireNet addresses the small ICBHI training set with data-efficient techniques for DNN classification.
1. INTRODUCTION
- Stethoscope-based diagnosis is low-cost and non-invasive, but requires trained professionals and is subject to inter-listener variability.The paper notes these limitations are exacerbated where expert medical professionals are scarce, and identifies automated analysis as a route to telemedicine monitoring.Source: §1 · p. 1
- Prior respiratory-sound DNNs use time-frequency inputs such as Mel-spectrograms, MFCC features, or S-transform spectrograms with CNN, RNN, or hybrid CNN-RNN models.The paper situates its simple CNN approach against earlier hand-crafted-feature and traditional machine-learning methods.Source: §1 · p. 1
- ICBHI contains only 6,898 breathing cycles, a small training set for data-hungry DNNs, motivating techniques to use the available data efficiently.The paper reports that a simple CNN architecture such as ResNet is adequate, contrasting with prior work using more complex architectures.Source: §1 · p. 2
- The paper proposes device-specific fine-tuning, concatenation-based augmentation, blank region clipping, and smart padding to address dataset characteristics that inhibit DNN training.The techniques are presented as orthogonal to network architecture and easy to incorporate into other networks.Source: §1 · p. 2
2. METHOD
RespireNet uses a ResNet34-based classifier and analyzes ICBHI dataset characteristics to address limited data, class imbalance, variable cycle lengths, blank spectrogram regions, and device imbalance. Its techniques include augmentation, smart padding, blank-region clipping, and device-specific fine-tuning.
2. METHOD
- The ICBHI respiratory sound dataset comprises 920 recordings from 126 patients, totaling 5.5 hours; experts annotate each breathing cycle as normal, crackle, wheeze, or both.Recordings come from four devices; data were collected at seven body locations per patient at hospitals in Portugal and Greece.Source: §2 · p. 2
- RespireNet downsamples audio to 4 kHz, filters and normalizes it, converts it to a Mel-spectrogram, and classifies it with ResNet34 and two 128-dimensional fully connected layers.The classifier uses softmax probabilities, dropout, and categorical cross-entropy loss.Source: §2 · p. 2
2.1. Efficient Dataset Utilization
- To address class imbalance, concatenation-based augmentation joins two randomly selected samples of the same class, improving classification accuracy for abnormal classes.The normal class accounts for 53% of samples; the approach is used alongside standard audio augmentation and a weighted random sampler.Source: §2.1 · p. 2, Figure 2 · p. 2
- Smart padding extends cycles using a neighboring same-class or normal cycle, or repeats the current cycle; a 7-second input window performed best.Cycle lengths range from 0.2 to 16.2 seconds, with a mean of 2.7 seconds; shorter windows clip samples, while very long windows cause repetition.Source: §2.1 · pp. 2–3
- Blank region clipping removes blank high-frequency spectrogram rows, while device-specific fine-tuning adapts a full-data model using each device's training subset.All Litt3200 samples had blank regions from 1500–2000 Hz, and AKGC417L contributed 63% of samples; the authors report fine-tuning particularly improves performance for underrepresented devices.Source: §2.1 · p. 3, Figure 3 · p. 3
3. EXPERIMENTS
RespireNet achieved state-of-the-art results on both ICBHI train-test divisions and the 2-class task. Ablations report gains from concatenation-based augmentation, smart padding, blank region clipping, and device-specific fine-tuning.
Evaluation setup
- The evaluation covers 4-class and normal-versus-anomalous classification, with the ICBHI score defined as the mean of sensitivity and specificity.The train and test sets are patient-wise disjoint for both divisions.Source: §3 · p. 3
Overall results
- 4% on the official 60-40 split and 2.2% on the 80-20 split over the respective state-of-the-art systems; the 2-class score reached 77%.The evaluation uses patient-wise disjoint splits.Source: §3 · p. 3, Table 1 · p. 4
Ablation analysis
- Concatenation-based augmentation raised the score from 62.2% to 66.8% after standard augmentation reached 66.2%, with abnormal-class sensitivity increasing by 1.5%.The authors attribute the sensitivity gain to improved abnormal-class generalization from generating novel abnormal samples.Source: §3 · pp. 3–4
- Smart padding improved base-model accuracy by 5% over zero-padding and 2% over reflection-based padding; a 7s input length performed optimally.The mean cycle length is 2.7s, so most inputs require padding.Source: §3 · p. 4
- Blank Region Clipping improved the base-model score of 66.2% by 0.5%; combined with the proposed augmentation, it achieved 67.1%, outperforming the current SOTA by 0.8%.Source: §3 · p. 4
- Device-specific fine-tuning improved the final ICBHI score by 1.4%, while devices with fewer samples had score increases of about 9%.Table 3 is identified as reporting the device-specific fine-tuning results.Source: §3 · p. 4, Table 3 · p. 4
4. RELATED WORK
Prior respiratory-sound work used deep learning and time-frequency audio representations, with reported performance across different tasks and dataset splits. RespireNet’s related-work discussion highlights that existing approaches did not focus on ICBHI dataset characteristics.
4. RELATED WORK
- Deep learning respiratory-sound classifiers had outperformed HMM-GMM and traditional machine-learning methods, commonly using time-frequency audio representations.The cited traditional methods include boosted decision trees and SVMs.Source: §4 · p. 4
- Kochetov et al. reported 65.7% on the 80-20 four-class split; Chen et al. reported 98% for three-class anomaly classification on a smaller ICBHI subset.The authors note that Kochetov et al. omitted noise-label-generation details, while Chen et al. used a 70-30 split.Source: §4 · p. 4
- Acharya and Basu reported 66.3% on four-class 80-20 evaluation, while Ma et al. reported 63.7% sensitivity; neither focused on ICBHI characteristics.The cited approaches use patient-specific tuning and a non-local ResNet block with mixup, respectively.Source: §4 · p. 4
5. CONCLUSION AND FUTURE WORK
The paper concludes that RespireNet and its data-use techniques achieve new state-of-the-art results on ICBHI’s 2-class and 4-class tasks. It identifies dataset size, device variation, and 80-20 split variability as boundaries on performance and comparison.
5. CONCLUSION AND FUTURE WORK
- RespireNet combines a simple CNN with four techniques for using small respiratory-sound datasets, achieving new ICBHI state-of-the-art results on both 2-class and 4-class tasks.The techniques are device-specific fine-tuning, concatenation-based augmentation, blank region clipping, and smart padding.Source: §5 · p. 4
- The techniques are described as orthogonal to network architecture and easy to incorporate into other frameworks.Source: §5 · p. 4
- The authors attribute the 4-class performance limit mainly to ICBHI’s small size and recording-device variation, and report variance across 80-20 splits.They recommend capturing a larger dataset while addressing the issues raised in the paper.Source: §5 · p. 4
6. SUPPLEMENTARY MATERIAL
The supplement describes the ICBHI dataset’s size and imbalances, then presents additional results on single-device training, blank-region clipping, and device-specific fine-tuning.
6.1. Dataset Details
- The dataset contains 920 recordings from 126 patients, totaling 5.5 hours, with cycle lengths from 0.2 to 16.2 seconds and four imbalanced classes.Class counts are 3,642 normal, 1,864 crackle, 886 wheeze, and 506 both.Source: §6.1 · p. 6
- 65% of samples have cycles shorter than 3 seconds, while 33% have cycles between 4 and 6 seconds.Source: Figure 4 · p. 6
- 64 Meditron patients averaged 22.9 cycles each, versus 32 AKGC417L patients averaging 136.4; 83% of crackle samples were from AKGC417L.The supplement identifies these distributions as a substantial skew across devices and patients.Source: §6.1 · p. 6, Table 4 · p. 6
6.2. Additional Results
- 53.0% overall score, 55.7% sensitivity, and 50.3% specificity are reported when training only on AKGC417L.The accompanying table reports test performance by device.Source: Table 6 · p. 6, §6.2 · p. 6
- After blank-region clipping, test-set attention maps focus more on the spectrogram’s bottom half rather than blank spaces, and performance improves.The visualization averages Grad-CAM++ maps for 1,370 test samples.Source: §6.2 · p. 6, Figure 5 · p. 7
- Figure 6 presents confusion matrices before and after device-specific fine-tuning.Source: §6.2 · p. 6, Figure 6 · p. 7