Integrating
nnU
‐Net Segmentation and Clinical–Radiomics for Multicenter
MRI
‐Based Assessment of Soft Tissue Sarcoma Grade and Ki‐67 Expression
Jinge Li, Yifeng Zhu, Honghai Chen, Lina Zhang, Xiangwen Li, Jie Zhou, Kai Zhang, Jie Huang, Xinyu Yang, Jiaye Zhang, Yushi Li, Wenjia Wang, Juan Tao, Shaowu Wang ABSTRACT
Background
Histological grade and Ki‐67 expression are prognostic risk factors in patients with soft tissue sarcoma (STS). These assessments require biopsy, which may be affected by tumor heterogeneity and is invasive.
Purpose
To develop an automated MRI‐based pipeline to assess STS grade and Ki‐67 expression.
Study Type
Retrospective.
Population
186 patients with pathological confirmation of STS (89 low‐grade, 97 high‐grade; 87 low Ki‐67 expression, 99 high Ki‐67 expression) across three hospitals, with 130 and 56 patients in the training and validation cohorts.
Field Strength/Sequence
1.5 T and 3.0 T/Fat‐suppressed T2‐weighted imaging, fat‐suppressed gadolinium‐enhanced T1‐weighted imaging, and diffusion‐weighted imaging.
Assessment
An automatic STS segmentation model was developed and compared with manual segmentations. Clinical‐imaging signature (CS) models were developed to distinguish STS grade and Ki‐67 expression using: (i) structural MRI (T1WI and T2WI) radiomics features, (ii) structural MRI and ADC radiomics features, and (iii) structural MRI and ADC radiomics features combined with clinical information and MRI semantic features. The diagnostic performance of radiologists (with 6, 8, and 38 years' experience) for assessing grade and Ki‐67 expression was evaluated with and without the assistance of the best‐performing models.
Statistical Tests
Dice coefficient, Cohen's κ and weighted κ , chi‐square test or Fisher's exact test, logistic regression analyses, decision curve analysis, area under the receiver operating characteristic curve (AUC), and DeLong's test. A p value < 0.05 was considered significant.
Results
The segmentation model achieved good segmentation performance (0.80 in extremity cases and 0.73 in trunk cases). LR and SVM CS models showed the best performance for grading and Ki‐67 assessment, respectively. (AUC in validation cohort: 0.846 and 0.742). Using the model significantly improved the diagnostic performance of the two more‐junior radiologists for grade (AUC: 0.720–0.832 and 0.735–0.835) and Ki‐67 expression (AUC: 0.665–0.717 and 0.659–0.741).
Data Conclusion
The CS model may assess STS grade and Ki‐67 expression and improve the diagnostic performance of less‐experienced radiologists.
Evidence Level
3.
Technical Efficacy
2.