Multi-Modal Ultrasound Lung Segmentation Using Noise-Resilient Deep Learning Architecture

Ultrasound imaging is widely used for lung diagnostics due to its non-invasive and cost-effective nature. However, challenges like speckle noise and low resolution make lung segmentation difficult. This paper presents a novel approach combining multi-modal learning, noise-resilient segmentation, and edge-device optimization to enable effective and accessible ultrasound lung segmentation. The method leverages feature transfer from high-resolution modalities like CT scans and employs a hybrid architecture with denoising autoencoders based on convolutional neural networks to enhance segmentation accuracy, even for low-quality ultrasound images. Lightweight architecture and quantization techniques optimize the system for real-time inference, making it suitable for portable ultrasound devices. Extensive experiments on benchmark datasets demonstrate superior segmentation accuracy and generalizability compared to state-of-the-art methods. This innovation supports point-of-care applications in resource-limited settings, advancing lung diagnostics for underserved communities.