Scholarlink

Pulmonary ultrasound (LUS) is a critical non-invasive tool for assessing lung aeration particularly in patients susceptible to atelectasis. However, LUS interpretation is operator-dependent and requires expertise. This study investigates the use of deep learning, specifically Convolutional Neural Networks (CNNs) with transfer learning, for automated classification of LUS aeration patterns (Normal A-lines, B1-lines, B2-lines, and Consolidation). We evaluated three CNN architectures (AlexNet, ResNet50, and MobileNet) on a dataset of 911 expert-labeled LUS images. Furthermore, we explored the impact of a Denoising Autoencoder on classification performance by preprocessing images to reduce noise. Our results demonstrate that MobileNet achieved the highest validation accuracy of 89.42%, outperforming AlexNet (88.32% accuracy) and ResNet50 (53.65% accuracy). When MobileNet was applied to noisy images reconstructed by the Denoising Autoencoder, it maintained a robust performance with 87.23% accuracy, indicating the potential of this combined approach for handling noisy or lower-quality images. This research and previous research provide evidence for the efficacy of transfer learning with efficient CNNs like MobileNet for LUS analysis. Future work includes expanding the training dataset, optimizing the autoencoder for improved noise reduction with minimal information loss, validating these models in diverse clinical settings, and exploring sequential models for dynamic LUS video analysis. These advancements are expected to provide robust automated tools to support clinical decision-making in lung aeration assessment.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Recognition of Lung Aeration Status Using Convolutional Neural Networks with Transfer Learning and Autoencoder-Based Noise Reduction

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Antonella Marini Brancato,
  • Candelaria Urroz Morales,
  • Diego Sebastian Comas,
  • Cecilia Acosta,
  • Gerardo Tusman,
  • Gustavo Javier Meschino

摘要

Pulmonary ultrasound (LUS) is a critical non-invasive tool for assessing lung aeration particularly in patients susceptible to atelectasis. However, LUS interpretation is operator-dependent and requires expertise. This study investigates the use of deep learning, specifically Convolutional Neural Networks (CNNs) with transfer learning, for automated classification of LUS aeration patterns (Normal A-lines, B1-lines, B2-lines, and Consolidation). We evaluated three CNN architectures (AlexNet, ResNet50, and MobileNet) on a dataset of 911 expert-labeled LUS images. Furthermore, we explored the impact of a Denoising Autoencoder on classification performance by preprocessing images to reduce noise. Our results demonstrate that MobileNet achieved the highest validation accuracy of 89.42%, outperforming AlexNet (88.32% accuracy) and ResNet50 (53.65% accuracy). When MobileNet was applied to noisy images reconstructed by the Denoising Autoencoder, it maintained a robust performance with 87.23% accuracy, indicating the potential of this combined approach for handling noisy or lower-quality images. This research and previous research provide evidence for the efficacy of transfer learning with efficient CNNs like MobileNet for LUS analysis. Future work includes expanding the training dataset, optimizing the autoencoder for improved noise reduction with minimal information loss, validating these models in diverse clinical settings, and exploring sequential models for dynamic LUS video analysis. These advancements are expected to provide robust automated tools to support clinical decision-making in lung aeration assessment.