Scholarlink

Background <p>Current MRI prostate cancer risk assessment methods focus mainly on detecting tumor lesions, ignoring the prostate gland macro-environment which may also impact disease progression. A generalized deep-learning model for prostate may help capture these gland-level characteristics through deep embeddings which can be used for a variety of downstream tasks. This study aims to assess whether U-Found, a generalized multiparametric (mp)MRI-based model, offers added value in predicting histopathological progression in active surveillance (AS) patients. The prostate macro-environment, captured in U-Found embeddings, is hypnotized to play a significant role in differentiating patients who progress to definitive treatment from those whose tumor is kept at bay.</p> Methods <p>U-Found was trained on a dataset comprising over 3000 mpMRIs from in-house and public sources using self-supervised learning. Axial slices were represented in a 128-dimentional space. The physical interpretation of the embeddings was explored by investigating images that are closest to the centroid of embeddings clusters. U-Found was tested on a downstream task: identifying cancer in an independent dataset (publicly available UCLA dataset, <i>n</i> = 1,151). To determine the added value of U-Found embeddings to clinical and intratumoral radiomics features, we compared models for predicting histopathological progression in 144 participants of a prospective AS trial. In addition, associations between U-Found embeddings and lesion- and prostate radiomics were investigated.</p> Results <p>Our findings suggest that U-Found captures key characteristics of the prostate gland’s macro-environment. U-Found successfully detected cancer in an independent UCLA dataset without being explicitly trained for lesion detection (AUC = 0.79). The prediction model incorporating a combination of clinical variables, mpMRI-derived intratumoral radiomics features and deep embeddings generated by U-Found achieved AUC = 0.86, outperforming models solely based on clinical and/or radiomics features. There were clear associations between U-Found embeddings and radiomics features.</p> Conclusions <p>U-Found was designed as a generalized self-supervised foundation model for prostate imaging, enabling the model to learn intrinsic imaging structures. We demonstrate that U-Found embeddings capture key features of the prostate macro-environment, which appear to contribute to disease progression, albeit to a lesser extent than tumor-specific imaging features.</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

MRI-based patient selection for active surveillance in prostate cancer using U-Found: a generalized deep learning model

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月19日

查看全文
  • Noah C. Lowry,
  • Adrian L. Breto,
  • Veronica Wallaengen,
  • Ahmad Algohary,
  • Nicolas Tapia-Stoll,
  • Sandra M. Gaston,
  • Nachiketh S. Prakash,
  • Pedro F. S. Freitas,
  • Oleksandr N. Kryvenko,
  • Patricia Castillo,
  • Joel Saltz,
  • Tahsin Kurc,
  • Chad R. Ritch,
  • Bruno Nahar,
  • Mark L. Gonzalgo,
  • Dipen J. Parekh,
  • Brandon Mahal,
  • Benjamin O. Spieler,
  • Alan Dal Pra,
  • Matthew C. Abramowitz,
  • Alan Pollack,
  • Sanoj Punnen,
  • Radka Stoyanova

摘要

Background

Current MRI prostate cancer risk assessment methods focus mainly on detecting tumor lesions, ignoring the prostate gland macro-environment which may also impact disease progression. A generalized deep-learning model for prostate may help capture these gland-level characteristics through deep embeddings which can be used for a variety of downstream tasks. This study aims to assess whether U-Found, a generalized multiparametric (mp)MRI-based model, offers added value in predicting histopathological progression in active surveillance (AS) patients. The prostate macro-environment, captured in U-Found embeddings, is hypnotized to play a significant role in differentiating patients who progress to definitive treatment from those whose tumor is kept at bay.

Methods

U-Found was trained on a dataset comprising over 3000 mpMRIs from in-house and public sources using self-supervised learning. Axial slices were represented in a 128-dimentional space. The physical interpretation of the embeddings was explored by investigating images that are closest to the centroid of embeddings clusters. U-Found was tested on a downstream task: identifying cancer in an independent dataset (publicly available UCLA dataset, n = 1,151). To determine the added value of U-Found embeddings to clinical and intratumoral radiomics features, we compared models for predicting histopathological progression in 144 participants of a prospective AS trial. In addition, associations between U-Found embeddings and lesion- and prostate radiomics were investigated.

Results

Our findings suggest that U-Found captures key characteristics of the prostate gland’s macro-environment. U-Found successfully detected cancer in an independent UCLA dataset without being explicitly trained for lesion detection (AUC = 0.79). The prediction model incorporating a combination of clinical variables, mpMRI-derived intratumoral radiomics features and deep embeddings generated by U-Found achieved AUC = 0.86, outperforming models solely based on clinical and/or radiomics features. There were clear associations between U-Found embeddings and radiomics features.

Conclusions

U-Found was designed as a generalized self-supervised foundation model for prostate imaging, enabling the model to learn intrinsic imaging structures. We demonstrate that U-Found embeddings capture key features of the prostate macro-environment, which appear to contribute to disease progression, albeit to a lesser extent than tumor-specific imaging features.