Deep Task-Based Beamforming and Channel Data Augmentations for Enhanced Ultrasound Imaging

TL;DR

This paper presents a deep learning framework for task-based ultrasound beamforming that integrates clinical tasks into the process, improving image quality and clinical relevance through novel architectures and data augmentations.

Contribution

It introduces two innovative deep learning-based beamforming methods that incorporate clinical tasks and channel data augmentations to enhance ultrasound image quality.

Findings

Channel data augmentations improve image quality significantly.

The CDCB approach outperforms traditional beamforming methods.

Proposed methods enhance clinical relevance of ultrasound images.

Abstract

This paper introduces a deep learning (DL)-based framework for task-based ultrasound (US) beamforming, aiming to enhance clinical outcomes by integrating specific clinical tasks directly into the beamforming process. Task-based beamforming optimizes the beamformer not only for image quality but also for performance on a particular clinical task, such as lesion classification. The proposed framework explores two approaches: (1) a Joint Beamformer and Classifier (JBC) that classifies the US images generated by the beamformer to provide feedback for image quality improvement; and (2) a Channel Data Classifier Beamformer (CDCB) that incorporates classification directly at the channel data representation within the beamformer's bottleneck layer. Additionally, we introduce channel data augmentations to address challenges posed by noisy and limited in-vivo data. Numerical evaluations demonstrate that training with channel data augmentations significantly improves image quality. The proposed methods were evaluated against conventional Delay-and-Sum (DAS) and Minimum Variance (MV) beamforming techniques, demonstrating superior performance in terms of both image contrast and clinical relevance. Among all methods, the CDCB approach achieves the best results, outperforming others in terms of image quality and clinical relevance. These approaches exhibit significant potential for improving clinical relevance and image quality in ultrasound imaging.