GraphULM: A Multi-Resolution CNN and GCN Framework for Ultrasound Localization Microscopy

Mohammad Sabih, Mohamed Khaled Almekkawy

Keywords: Ultrasound Localization Microscopy, Multi-Resolution CNN, Graph Convolu- tional Network, Attention Mechanism, Medical Imaging

Abstract:

Ultrasound Localization Microscopy (ULM) is a prominent technique in medical imaging, widely applied to enhance super-resolution, particularly in in-vivo settings. The process of localization, followed by tracking, poses a significant challenge in ULM due to its intricacy and complexity. High microbubble densities intensify these challenges, thereby diminishing the performance of traditional methods
and certain deep learning algorithms in achieving precise localization. We present GraphULM, a novel and computationally efficient architecture that combines a Multi-Resolution Convolutional Neural Network (MRCNN) with a Graph Convolutional Network (GCN) to enhance localization efficacy in ULM. To develop an optimal training dataset, synthetically generated data is pre-combined with in-vivo b-mode samples, resulting in a perfect dataset that enhanced the generalization capability of our model. Experimental evaluations in in-vivo demonstrate the model’s high performance, reporting a localization precision of 21.9 micro meter, and a Jaccard index of 0.75, at a microbubble density of 2 MB/mm2, underscoring the model’s robustness. Additionally, our Frequency Ring Correlation (FRC) analysis reveals a remarkable resolution of 5.62 micro meter. The model operates at three times the speed of traditional pipelines, establishing its suitability for rapid ULM applications.

Primary Subject Area: Image Acquisition and Reconstruction

Secondary Subject Area: Image Synthesis

Registration Requirement: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 372