Scholarlink

Meningitis is a life-threatening disease, resulting in severe neurological damage or death if not treated in time. The standard diagnostic method is an invasive lumbar puncture (LP), not exempt of complications and not always feasible, especially in resource-poor settings. To overcome these challenges, we developed Neosonics®, a novel non-invasive high-resolution ultrasound (HRUS) technology that detects backscatter signals from white blood cells (WBCs) in cerebrospinal fluid (CSF) below the infant fontanel. Using deep learning (DL) for image analysis, this technology classifies patients based on WBC levels, providing a rapid, non-invasive screening method for infant meningitis and with this, limiting indications for LPs for positive cases only, if not contraindicated. A convolutional neural network was trained and validated using CSF HRUS patient data gathered from cohorts from three different countries, Mozambique, Spain and Morocco. Then, a soft voting ensemble-learning technique was applied to give a classification on a patient-level. The DL model showed on a single-image level a sensitivity (SE) and specificity (SP) of 71.1% and 87.0% for the Mozambican cohort, SE = 73.3% and SP = 75.5% for the Spanish cohort, and SE = 72.0% and SP = 88.7% for the Moroccan cohort. On a patient-level, the overall SE and SP was 94.4% and 94.8%, respectively. We demonstrate in this study the significance of combining HRUS and DL for the non-invasive detection of infant meningitis, offering an efficient, cost-effective solution suitable also for limited resource settings and remote areas.

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

Deep Learning Based Non-invasive Meningitis Screening Using High-Resolution Ultrasound in Neonates and Infants from Mozambique, Spain and Morocco

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Beatrice M. Jobst,
  • Francesc Carandell,
  • Sara Ajanovic,
  • Hassan Sial,
  • Javier Jiménez,
  • Rita Quesada,
  • Fabião Santos,
  • Manuela Lopez-Azorín,
  • Eva Valverde,
  • Marta Ybarra,
  • M. Carmen Bravo,
  • David Muñoz,
  • Thais Agut,
  • Barbara Salas,
  • Nuria Carreras,
  • Ana Alarcón,
  • Martín Iriondo,
  • Carles Luaces,
  • Muhammad Sidat,
  • Mastalina Zandamela,
  • Paula Rodrigues,
  • Luzidina Martins,
  • Uneisse Cassia,
  • Justina Bramugy,
  • Anelsio Cossa,
  • Campos Mucasse,
  • W. Chris Buck,
  • Sara Arias,
  • Chaymae El Abbass,
  • Houssain Tligui,
  • Amina Barkat,
  • Najat Amalik,
  • Imane Zizi,
  • Alberto Ibáñez,
  • Montserrat Parrilla,
  • Luis Elvira,
  • Cristina Calvo,
  • Adelina Pellicer,
  • Fernando Cabañas,
  • Quique Bassat,
  • Paula Petrone

摘要

Meningitis is a life-threatening disease, resulting in severe neurological damage or death if not treated in time. The standard diagnostic method is an invasive lumbar puncture (LP), not exempt of complications and not always feasible, especially in resource-poor settings. To overcome these challenges, we developed Neosonics®, a novel non-invasive high-resolution ultrasound (HRUS) technology that detects backscatter signals from white blood cells (WBCs) in cerebrospinal fluid (CSF) below the infant fontanel. Using deep learning (DL) for image analysis, this technology classifies patients based on WBC levels, providing a rapid, non-invasive screening method for infant meningitis and with this, limiting indications for LPs for positive cases only, if not contraindicated. A convolutional neural network was trained and validated using CSF HRUS patient data gathered from cohorts from three different countries, Mozambique, Spain and Morocco. Then, a soft voting ensemble-learning technique was applied to give a classification on a patient-level. The DL model showed on a single-image level a sensitivity (SE) and specificity (SP) of 71.1% and 87.0% for the Mozambican cohort, SE = 73.3% and SP = 75.5% for the Spanish cohort, and SE = 72.0% and SP = 88.7% for the Moroccan cohort. On a patient-level, the overall SE and SP was 94.4% and 94.8%, respectively. We demonstrate in this study the significance of combining HRUS and DL for the non-invasive detection of infant meningitis, offering an efficient, cost-effective solution suitable also for limited resource settings and remote areas.