Scholarlink

Purpose <p>A deep learning algorithm for contrast amplification in brain MRI, trained exclusively on adult data, was tested for cross-population generalization to pediatric patients, including subjects aged 0–2&#xa0;years.</p> Methods <p>A retrospective monocentric dataset (n = 22 cases) comprising pediatric patients (0–17&#xa0;years old) diagnosed with various brain tumors was used to evaluate the algorithm, which takes T1-weighted pre- and standard post-contrast images as input and generates an output image with amplified contrast, further post-processed with an HDR algorithm. Quantitative comparisons between standard and amplified images were performed using contrast-to-noise ratio (CNR), contrast enhancement percentage (CEP), and lesion-to-background ratio (LBR). Three neuroradiologists performed qualitative assessment using a 4-point Likert scale, focusing on lesion contrast and delineation. Anatomical similarity was assessed using SSIM and log-Jacobian range. Statistical significance was evaluated using two-tailed paired t-tests.</p> Results <p>Compared to standard-dose images, contrast-amplified images showed significantly higher values for CNR (+ 186.5%), LBR (+ 61.9%), and CEP (+ 110.4%). Qualitative assessments demonstrated comparable lesion visualization, with improvements observed in selected cases. Reader 1 preferred the contrast-amplified image in 12 of 22 cases (54.5%), reader 2 favored it in 18 of 22 cases (81.8%) and reader 3 in 13/22 cases (59.1%). One reader reported improved overall image quality (mean score: 3.95 vs. 3.73). The average SSIM between amplified and standard-dose images was 0.98, and any significant anatomical differences were highlighted by the log-Jacobian range (<i>p-</i>value = 0.556).</p> Conclusion <p>An algorithm for contrast amplification based on deep learning, trained with adult data, significantly enhances quantitative contrast metrics in images from pediatric patients. It is preferred over standard-dose images in the majority of cases when used for pediatric brain MRI, indicating its promising application for cross-population applicability.</p>

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 for contrast-enhanced MRI in pediatric brain imaging

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月4日

查看全文
  • Anna Macula,
  • Giovanni Morana,
  • Fiorenza Coppola,
  • Andrea Garnero,
  • Andrea Rossi,
  • Alberto Fringuello Mingo,
  • Stefano Tambalo,
  • Giovanni Valbusa,
  • Fabio Tedoldi,
  • Sonia Colombo Serra,
  • Angelo Bifone

摘要

Purpose

A deep learning algorithm for contrast amplification in brain MRI, trained exclusively on adult data, was tested for cross-population generalization to pediatric patients, including subjects aged 0–2 years.

Methods

A retrospective monocentric dataset (n = 22 cases) comprising pediatric patients (0–17 years old) diagnosed with various brain tumors was used to evaluate the algorithm, which takes T1-weighted pre- and standard post-contrast images as input and generates an output image with amplified contrast, further post-processed with an HDR algorithm. Quantitative comparisons between standard and amplified images were performed using contrast-to-noise ratio (CNR), contrast enhancement percentage (CEP), and lesion-to-background ratio (LBR). Three neuroradiologists performed qualitative assessment using a 4-point Likert scale, focusing on lesion contrast and delineation. Anatomical similarity was assessed using SSIM and log-Jacobian range. Statistical significance was evaluated using two-tailed paired t-tests.

Results

Compared to standard-dose images, contrast-amplified images showed significantly higher values for CNR (+ 186.5%), LBR (+ 61.9%), and CEP (+ 110.4%). Qualitative assessments demonstrated comparable lesion visualization, with improvements observed in selected cases. Reader 1 preferred the contrast-amplified image in 12 of 22 cases (54.5%), reader 2 favored it in 18 of 22 cases (81.8%) and reader 3 in 13/22 cases (59.1%). One reader reported improved overall image quality (mean score: 3.95 vs. 3.73). The average SSIM between amplified and standard-dose images was 0.98, and any significant anatomical differences were highlighted by the log-Jacobian range (p-value = 0.556).

Conclusion

An algorithm for contrast amplification based on deep learning, trained with adult data, significantly enhances quantitative contrast metrics in images from pediatric patients. It is preferred over standard-dose images in the majority of cases when used for pediatric brain MRI, indicating its promising application for cross-population applicability.