This is the implementation of the work CycleGuardian:A Framework for Automatic RespiratorySound classification Based on Improved Deep clustering and Contrastive Learning.
If you like it give it star.
python3.8.18
#
# Name Version Build Channel
pytorch-wavelets 1.3.0 pypi_0 pypi
torch 2.0.1 pypi_0 pypi
torch-geometric 2.5.3 pypi_0 pypi
torchaudio 2.0.2 pypi_0 pypi
torchvision 0.15.2 pypi_0 pypi
Dependencies are listed in torch2.0.1_requirements.txt.
These are all easily installable via, e.g., pip install numpy or conda install.
It is recommended to use Anaconda to create a virtual environment and setup the dependencies .
- The dataset used is the ICBHI Respiratory Challenge 2017 dataset.
- To find out more details about the dataset please visit the official ICBHI Challenge webpage.
- Download from above and place the dataset in the
datafolder.
- We follow both the official
60-40train-test split as well as the80-20split.
Change the num_worker to suit your machine; also, you may change the batch size, but it will affect the final score.
python train_lab8_6_3CycleGuardian_specAug.py --data_dir ./data/ICBHI_final_database --dataset_split_file ./data/patient_trainTest6_4.txt --model_path ./models_out --lr_h 0.001 --lr_l 0.001 --batch_size 128 --num_worker 15 --start_epochs 0 --epochs 600
| Method | Architecture/ Pretrained | ( S_p ) | ( S_e ) | Score | Size |
|---|---|---|---|---|---|
| 4 Class | |||||
| Path-level [^song2023patch] | CNN6 / IN | 68.59% | 48.25% | 58.42% | 20MB+ |
| RespirNet [^gairola2021respirenet] | Resnet34 / IN | 72.30% | 40.10% | 56.20% | 83MB+ |
| Domain transfor [^wang2022domain] | Resnet-st / IN | 70.40% | 40.20% | 55.30% | 50MB+ |
| Late fusion [^pham2022ensemble] | Ince-v3+VGG14 / IN | 85.60% | 30.00% | 57.30% | 610MB+ |
| Co-tuing [^nguyen2022lung] | Resnet50 / IN | 79.34% | 37.24% | 58.29% | 97MB+ |
| SCL [^moummad2023pretraining] | CNN6 / AS | 76.93% | 39.15% | 58.04% | 20MB+ |
| Patch mix [^bae2023patch] | AST / IN+AS | 81.66% | 43.07% | 62.37% | 380MB+ |
| OFGST [^wang2024ofgst] | Swin Transformer / IN | 71.56% | 40.53% | 56.05% | 410MB+ |
| Stethoscope [^kim2024stethoscope] | AST / IN+AS | 79.87% | 43.55% | 61.71% | 380MB+ |
| RepAug [^kim2024repaugment] | AST / IN+AS | 82.47% | 40.55% | 61.51% | 380MB+ |
| Lungbrn [^ma2019lungbrn] | bi-Resent / - | 69.20% | 31.12% | 50.16% | 80MB+ |
| SE+SA [^yang2020adventitious] | Resnet18 / - | 81.25% | 17.84% | 49.55% | 44MB+ |
| LungRN+NL [^ma2020lungrn+] | Resnet-NL / - | 63.20% | 41.32% | 52.26% | 40MB+ |
| Cnn+MOE [^pham2021cnn] | C-DNN / - | 72.40% | 21.50% | 47.00% | 30MB+ |
| LungAttn [^li2021lungattn] | ResNet-Atten / - | 71.44% | 36.36% | 53.90% | 60MB+ |
| ARSC-net [^xu2021arsc] | bi-Resnet-Att / - | 67.13% | 46.38% | 56.76% | 80MB+ |
| Prototype [^ren2022prototype] | Cnn8-pt / - | 72.96% | 27.78% | 50.37% | 30MB+ |
| Adversarial [^chang2022example] | Cnn8-dilated / - | 69.92% | 35.85% | 52.89% | 30MB+ |
| Patch mix [^bae2023patch] | ASTransformer / - | 72.61% | 30.58% | 49.60% | 380MB+ |
| Ours(cos) | CycleGuardian / - | 88.92% | 33.33% | 61.12% | 27MB |
| Ours(soft cos) | CycleGuardian / - | 82.06% | 44.47% | 63.26% | 38MB |
| 2 Class | |||||
| Co-tuing [^nguyen2022lung] | Resnet50 / IN | 79.34% | 50.04% | 64.74% | 97MB+ |
| Patch mix [^bae2023patch] | ASTransformer / IN +AS | 81.66% | 55.77% | 68.71% | 380MB+ |
| Cnn+MOE [^pham2021cnn] | C-DNN / - | 72.40% | 37.50% | 54.10% | 30MB+ |
| Ours | CycleGuardian / - | 73.19% | 59.57% | 66.38% | 27MB |
| Ours(soft cos) | CycleGuardian / - | 71.93% | 62.77% | 67.35% | 38MB |
train loss
==================================================
epoch 599 iter 33/34 Train Total loss: 0.5795828700065613
Train Accuracy: 0.999061473486626
Classwise_Losses Normal: 0.002527494459287292, Crackle: 0.0021995086581145045, Wheeze: 0.005279731647609176, Both: 0.008709785856947726
-----show the training info -----------
*************The official Metrics******************
The frac format Sp: 0.9995196925552585, Se: 0.998623853073584, Score: 0.9990717728144213
The int format S_p: 99.95196925552105, S_e: 99.86238530735383, Score: 99.90717728143744
Normal: 0.9995196925552585, Crackle: 0.9991980753007862, Wheeze: 0.9980506820666568, Both: 0.9976190473815194
testing......
Val Accuracy by fusion result : 0.6181691611555865
epoch 599, Validation BCE loss: 2.727804660797119
Classwise_Losses Normal: 1.5925593402063463, Crackle: 3.354416726221597, Wheeze: 6.381649032856236, Both: 3.004433798120376
-----show the validation info -----------
*************The official Metrics******************
The frac format Sp: 0.7325361862806796, Se: 0.4766355139444493, Score: 0.6045858501125645
The int format S_p: 73.25361862806335, S_e: 47.663551394441214, Score: 60.45858501125228
Normal: 0.7325361862806796, Crackle: 0.5457227138643068, Wheeze: 0.27999999993, Both: 0.6310679608587049
*************The Helper Metrics******************
Confusion Matrix
[[1164 295 54 76]
[ 283 370 12 13]
[ 203 48 112 37]
[ 29 14 33 130]]
Classwise Scores -->: [0.73253619 0.54572271 0.28 0.63106796]
Accuracy Score ---> : 0.6181691611555865
Micro F1 Score: 0.6181691611555865
weighted Precision: 0.6138528596527835
weighted Recall: 0.6181691611555865
BEST ACCURACY TILL NOW 0.652627915071354
training process: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 34/34 [03:41<00:00, 6.50s/it]
best_Se0.44470404977496825 best_Sp0.8206419131529263 best_Sc0.6326729814639473 best_Acc0.652627915071354
ds combine best_Confusion_matrix:
[[0.82064191 0.12083071 0.03398364 0.02454374]
[0.49115044 0.48082596 0.01917404 0.00884956]
[0.5825 0.0825 0.3 0.035 ]
[0.19417476 0.03398058 0.16504854 0.60679612]]
@misc{chu2025CycleGuardian,
title={ycleGuardian: A Framework for Automatic RespiratorySound classification Based on Improved Deep clustering andContrastive Learning},
author={Yun Chu,Qiuhao Wang, Enze Zhou, Ling Fu, Qian Liu, Gang Zheng},
year={2025},
eprint={2025.00734},
archivePrefix={arXiv},
primaryClass={cs.AI}
}