Scholarlink

Background <p>Pancreatic neuroendocrine tumor (PNET) behavior depends on tumor grade and genomics. Radiomics may identify these factors noninvasively. This study developed an automated pipeline from segmentation to radiomics modeling to preoperatively predict tumor grade.</p> Patients and Methods <p>Patients resected from 2003 to 2021 with adequate preoperative arterial phase computed tomography (CT) scans were divided into training and test cohorts. The training cohort underwent manual pancreas and tumor region segmentation to train an auto-segmentation model; radiomic features extracted from tumor regions were used to develop a radiomics model for grade prediction (I versus II/III), which was evaluated in the automatically segmented test cohort. Associations between radiomic and genomic features were assessed.</p> Results <p>In total, 182 patients were divided into training (<i>n </i>= 140) and test (<i>n </i>= 42) cohorts. Grade I and II/III lesions were in 113 (62%) and 69 (38%) patients, respectively. Median tumor size was 24 mm (6, 200). The auto-segmentation model segmented tumor regions in 90% (38) of the test cohort. In this group (<i>n </i>= 38), the radiomics model produced receiver operating characteristic curve (AUC) of 0.85 (0.73, 0.96). Sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were 0.88 (0.78,0.98), 0.62 (0.46, 0.77), 0.65 (0.50, 0.80), and 0.87 (0.76,0.97), respectively. DAXX loss was associated with three radiomic features, and ATRX loss with one. In &lt; 2&#xa0;cm lesions, auto-segmentation was successful in 75% (9/12) of the test cohort, with accurate grade prediction in 67% (6/9) of cases.</p> Conclusions <p>The auto-segmentation model correctly identified tumor regions, and the radiomics model accurately predicted grade; new associations between genomic and radiomic features were identified. This automated pipeline can incorporate a radiomics model into preoperative PNET decision-making.</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

Automated Radiomics Model for Preoperative Pancreatic Neuroendocrine Tumor Grade Prediction

  • 期刊论文

  • 发表日期：

    2026年3月2日

查看全文
  • Pratik Chandra,
  • Hadi Ghahremannezhad,
  • Naaz Nasar,
  • Mithat Gonen,
  • Richard K. G. Do,
  • Carlie Sigel,
  • Alessandra Pulvirenti,
  • Kevin Soares,
  • Vinod Balachandran,
  • Jeffrey Drebin,
  • Michael D’Angelica,
  • T. Peter Kingham,
  • William R. Jarnagin,
  • Jayasree Chakraborty,
  • Alice C. Wei

摘要

Background

Pancreatic neuroendocrine tumor (PNET) behavior depends on tumor grade and genomics. Radiomics may identify these factors noninvasively. This study developed an automated pipeline from segmentation to radiomics modeling to preoperatively predict tumor grade.

Patients and Methods

Patients resected from 2003 to 2021 with adequate preoperative arterial phase computed tomography (CT) scans were divided into training and test cohorts. The training cohort underwent manual pancreas and tumor region segmentation to train an auto-segmentation model; radiomic features extracted from tumor regions were used to develop a radiomics model for grade prediction (I versus II/III), which was evaluated in the automatically segmented test cohort. Associations between radiomic and genomic features were assessed.

Results

In total, 182 patients were divided into training (n = 140) and test (n = 42) cohorts. Grade I and II/III lesions were in 113 (62%) and 69 (38%) patients, respectively. Median tumor size was 24 mm (6, 200). The auto-segmentation model segmented tumor regions in 90% (38) of the test cohort. In this group (n = 38), the radiomics model produced receiver operating characteristic curve (AUC) of 0.85 (0.73, 0.96). Sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were 0.88 (0.78,0.98), 0.62 (0.46, 0.77), 0.65 (0.50, 0.80), and 0.87 (0.76,0.97), respectively. DAXX loss was associated with three radiomic features, and ATRX loss with one. In < 2 cm lesions, auto-segmentation was successful in 75% (9/12) of the test cohort, with accurate grade prediction in 67% (6/9) of cases.

Conclusions

The auto-segmentation model correctly identified tumor regions, and the radiomics model accurately predicted grade; new associations between genomic and radiomic features were identified. This automated pipeline can incorporate a radiomics model into preoperative PNET decision-making.