Crackle Detection In Lung Sounds Using Transfer Learning And Multi-Input Convolutional Neural Networks.

Area of Science:

• Computational Intelligence and Machine Learning
• Medical Signal Processing
• Respiratory Medicine

Background:

• Large public lung sound databases exist for training diagnostic algorithms.
• Developing algorithms for small, non-public datasets with varying recording conditions presents challenges.
• Transfer learning offers a potential solution to bridge data domain discrepancies.

Purpose of the Study:

• To apply transfer learning to address recording setup mismatches in lung sound analysis.
• To improve crackle detection in lung sounds using knowledge transfer between datasets.
• To classify crackles and normal lung sounds on a self-collected dataset.

Main Methods:

• A single-input convolutional neural network (CNN) was pre-trained on the ICBHI 2017 lung sound database.
• Log-mel spectrogram features of respiratory cycles were utilized.
• A multi-input CNN model, sharing architecture for respiratory cycles and phases, was fine-tuned on a target dataset using transfer learning.

Main Results:

• The multi-input CNN model combined with transfer learning demonstrated significant performance improvements.
• An absolute increase of 9.84% in F-score was achieved on the target domain for crackle detection.
• The approach effectively transferred knowledge from a large public dataset to a smaller, specific dataset.

Conclusions:

• Transfer learning is a viable strategy to overcome domain mismatch issues in lung sound analysis.
• The proposed multi-input CNN model enhances crackle detection performance on challenging datasets.
• This method holds promise for developing more robust and accurate automated lung sound diagnosis systems.