MRI-Based Machine Learning Radiomics for Preoperative Assessment of Human Epidermal Growth Factor Receptor 2 Status in Urothelial Bladder Carcinoma

Ruixi Yu; Lingkai Cai; Yuxi Gong; Xueying Sun; Kai Li; Qiang Cao; Xiao Yang; Qiang Lu

doi:10.1002/jmri.29342

MRI-Based Machine Learning Radiomics for Preoperative Assessment of Human Epidermal Growth Factor Receptor 2 Status in Urothelial Bladder Carcinoma

J Magn Reson Imaging. 2024 Mar 8. doi: 10.1002/jmri.29342. Online ahead of print.

Authors

Ruixi Yu¹, Lingkai Cai^{1

2}, Yuxi Gong³, Xueying Sun⁴, Kai Li¹, Qiang Cao¹, Xiao Yang¹, Qiang Lu¹

Affiliations

¹ Department of Urology, The First Affiliated Hospital of Nanjing Medical University, Nanjing, China.
² Department of Urology, Wuxi Medical Center, Nanjing Medical University, Wuxi, China.
³ Department of Pathology, The First Affiliated Hospital of Nanjing Medical University, Nanjing, China.
⁴ Department of Radiology, The First Affiliated Hospital of Nanjing Medical University, Nanjing, China.

PMID: 38456745
DOI: 10.1002/jmri.29342

Abstract

Background: The human epidermal growth factor receptor 2 (HER2) has recently emerged as hotspot in targeted therapy for urothelial bladder cancer (UBC). The HER2 status is mainly identified by immunohistochemistry (IHC), preoperative and noninvasive methods for determining HER2 status in UBC remain in searching.

Purposes: To investigate whether radiomics features extracted from MRI using machine learning algorithms can noninvasively evaluate the HER2 status in UBC.

Study type: Retrospective.

Population: One hundred ninety-five patients (age: 68.7 ± 10.5 years) with 14.3% females from January 2019 to May 2023 were divided into training (N = 156) and validation (N = 39) cohorts, and 43 patients (age: 67.1 ± 13.1 years) with 13.9% females from June 2023 to January 2024 constituted the test cohort (N = 43).

Field strength/sequence: 3 T, T2-weighted imaging (turbo spin-echo), diffusion-weighted imaging (breathing-free spin echo).

Assessment: The HER2 status were assessed by IHC. Radiomics features were extracted from MRI images. Pearson correlation coefficient and the least absolute shrinkage and selection operator (LASSO) were applied for feature selection, and six machine learning models were established with optimal features to identify the HER2 status in UBC.

Statistical tests: Mann-Whitney U-test, chi-square test, LASSO algorithm, receiver operating characteristic analysis, and DeLong test.

Results: Three thousand forty-five radiomics features were extracted from each lesion, and 22 features were retained for analysis. The Support Vector Machine model demonstrated the best performance, with an AUC of 0.929 (95% CI: 0.888-0.970) and accuracy of 0.859 in the training cohort, AUC of 0.886 (95% CI: 0.780-0.993) and accuracy of 0.846 in the validation cohort, and AUC of 0.712 (95% CI: 0.535-0.889) and accuracy of 0.744 in the test cohort.

Data conclusion: MRI-based radiomics features combining machine learning algorithm provide a promising approach to assess HER2 status in UBC noninvasively and preoperatively.

Evidence level: 2 TECHNICAL EFFICACY: Stage 3.

Keywords: HER2; MRI; machine learning; radiomics; urinary bladder neoplasm.

Abstract

Grants and funding