Machine learning-based pathomics signature of histology slides as a novel prognostic indicator in primary central nervous system lymphoma

Ling Duan; Yongqi He; Wenhui Guo; Yanru Du; Shuo Yin; Shoubo Yang; Gehong Dong; Wenbin Li; Feng Chen

doi:10.1007/s11060-024-04665-8

Machine learning-based pathomics signature of histology slides as a novel prognostic indicator in primary central nervous system lymphoma

J Neurooncol. 2024 Apr 1. doi: 10.1007/s11060-024-04665-8. Online ahead of print.

Authors

Ling Duan¹, Yongqi He¹, Wenhui Guo¹, Yanru Du², Shuo Yin¹, Shoubo Yang¹, Gehong Dong³, Wenbin Li⁴, Feng Chen⁵

Affiliations

¹ Department of Neuro-Oncology, Cancer Center, Beijing Tiantan Hospital, Capital Medical University, No.119 West Nansihuan Road, Beijing, 100070, China.
² Department of Pathology, Beijing Tiantan Hospital, Capital Medical University, No.119 West Nansihuan Road, Beijing, 100070, China.
³ Department of Pathology, Beijing Tiantan Hospital, Capital Medical University, No.119 West Nansihuan Road, Beijing, 100070, China. 13520157603@126.com.
⁴ Department of Neuro-Oncology, Cancer Center, Beijing Tiantan Hospital, Capital Medical University, No.119 West Nansihuan Road, Beijing, 100070, China. liwenbin@ccmu.edu.cn.
⁵ Department of Neuro-Oncology, Cancer Center, Beijing Tiantan Hospital, Capital Medical University, No.119 West Nansihuan Road, Beijing, 100070, China. chenfeng406@sina.com.

PMID: 38557926
DOI: 10.1007/s11060-024-04665-8

Abstract

Purpose: To develop and validate a pathomics signature for predicting the outcomes of Primary Central Nervous System Lymphoma (PCNSL).

Methods: In this study, 132 whole-slide images (WSIs) of 114 patients with PCNSL were enrolled. Quantitative features of hematoxylin and eosin (H&E) stained slides were extracted using CellProfiler. A pathomics signature was established and validated. Cox regression analysis, receiver operating characteristic (ROC) curves, Calibration, decision curve analysis (DCA), and net reclassification improvement (NRI) were performed to assess the significance and performance.

Results: In total, 802 features were extracted using a fully automated pipeline. Six machine-learning classifiers demonstrated high accuracy in distinguishing malignant neoplasms. The pathomics signature remained a significant factor of overall survival (OS) and progression-free survival (PFS) in the training cohort (OS: HR 7.423, p < 0.001; PFS: HR 2.143, p = 0.022) and independent validation cohort (OS: HR 4.204, p = 0.017; PFS: HR 3.243, p = 0.005). A significantly lower response rate to initial treatment was found in high Path-score group (19/35, 54.29%) as compared to patients in the low Path-score group (16/70, 22.86%; p < 0.001). The DCA and NRI analyses confirmed that the nomogram showed incremental performance compared with existing models. The ROC curve demonstrated a relatively sensitive and specific profile for the nomogram (1-, 2-, and 3-year AUC = 0.862, 0.932, and 0.927, respectively).

Conclusion: As a novel, non-invasive, and convenient approach, the newly developed pathomics signature is a powerful predictor of OS and PFS in PCNSL and might be a potential predictive indicator for therapeutic response.

Keywords: Histopathological images; Pathomics; Primary central nervous system lymphoma; Prognosis.