WFUMB Ultrasound Open (Dec 2024)

Validation of a deep-learning modular prototype to guide novices to acquire diagnostic ultrasound images from urinary system

  • Silvia Ossaba,
  • Áurea Diez,
  • Milagros Marti,
  • María Luz Parra-Gordo,
  • Rodrigo Alonso-Gonzalez,
  • Rebeca Tenajas,
  • Gonzalo Garzón

Journal volume & issue
Vol. 2, no. 2
p. 100049

Abstract

Read online

Importance: Artificial intelligence (AI) application in guiding the acquisition of ultrasonography images represents a pioneering field of research. A new developed hybrid deep-learning (DL) algorithm, trained on more than high quality 60.000 curated and labelled reference images (distilled from a set of more than 600.000 abdominal ultrasound images) from La Paz Hospital, can provide real-time prescriptive guidance for novice operators to obtain standard planes images of the target organs. Objective: This study aims to evaluate the capability of novice users to acquire diagnostic-quality abdominal ultrasound images of the urinary system using the deep-learning (DL)-based guiding research prototype provided by GMV. Design: Setting, and Participants: This prospective diagnostic study was conducted within the facilities of an academic hospital. A cohort of 24 technically-oriented volunteers, lacking prior knowledge in anatomy or medicine and without experience in conducting ultrasound examinations, was recruited. After a brief training session focused on various organs, each pair of volunteers performed scans of each other's urinary system, exclusively guided by AI support. These scans were subsequently repeated by experienced sonographers using identical ultrasound equipment but without AI assistance. Four radiologists, each with decades of experience, independently and blindly assessed the quality of each acquisition. Results: Over 90.6 % of the images scanned by volunteers were identified as valuable clinical picture, using only an AI-based guidance system. This is nearly comparable to the results achieved by experienced radiologists, who attained a 98.6 % success rate. Conclusions: This deep-learning (DL) prototype enables novices lacking experience in ultrasonography to acquire diagnostic ultrasound images suitable for subsequent expert evaluation. The modular prototype works on a large range of ultrasound device models and vendors. This advancement has the potential to extend the application of ultrasound beyond traditional clinical environments, particularly in situations requiring immediate anatomical and functional interrogation, as well as in resource-limited settings.

Keywords