Classifying subtypes and predicting survival of renal cell carcinoma using histopathology image-based deep learning

Fei Wang, Yuanzhe Geng, Ting Wang, Ke Zhao, Bin Xu, Dachuan Zhang, Hongzhou Cai, Qinbo Yuan, Hang Lü, Yue Zhang, Li Xu, Yangyang Sun, Hang Gong, Raphael N. Alolga, Xiangshan Fan, Gaoxiang Ma, Lian‐Wen QiResearch Square 2021 · DOI

Abstract

Classifying histopathological subtypes and predicting survival of renal cell carcinoma (RCC) patients are critical steps towards treatment. In this work, we first proposed a deep learning method involving patch-based segmentation, intelligent feature extraction and heatmap visualization for classifying RCC into clear cell RCC, papillary RCC, chromophobe RCC, and adjacent benign tissue. This algorithm was trained and validated using 2,374,446 patches, 6,340 whole-slide images, 2,399 patients from The Cancer Genome Atlas and 6 independent centers. The classifiers provided areas under the curves of 0.979 to 0.996 in the internal phase, and 0.914 to 0.995 in the 6-center external phase. Furthermore, a modified deep learning approach comprising automated detection of regions of interest, patch-level learning, and morphological features-based risk scoring was developed for predicting survival of clear cell RCC patients. The prognostication model provided a hazard ratio for poor versus good prognosis of 2.63 [95% confidence interval (CI) 1.53–4.50, P = 4.35e-4] in the testing set, and 2.57 [95% CI 1.43–4.64, P = 1.68e-3] in an independent validation set using multivariable analyses. In conclusion, the developed histopathology image-based deep learning frameworks have the clinical potential to assist pathologists in systematically evaluating histological information of RCC patients.

Introduction

Renal cell carcinoma (RCC) is not a single disease but a collection of distinct subtypes—primarily clear cell (ccRCC), papillary (pRCC), and chromophobe (chRCC)—each requiring different treatment strategies. Currently, the "gold standard" for diagnosis is the manual review of histopathological slides by a pathologist.

However, this process is time-consuming, inherently subjective, and prone to inter-observer variability. Furthermore, predicting a patient’s prognosis (survival) is a major challenge due to the extreme heterogeneity of these tumors. Existing manual grading systems, like the Fuhrman grade, often fail to provide precise survival predictions across different stages of the disease. This paper presents an automated, standardized framework using deep learning to address these critical clinical needs.

Methods and Materials

The researchers developed two distinct but related deep learning pipelines to process whole-slide images (WSIs):

  1. The Classifier Model: Using an Inception V3 architecture, the team implemented a patch-based approach. The model first performs a binary classification to distinguish tumor from normal tissue. If a tumor is detected, it proceeds to a three-way classification to identify the specific RCC subtype. Heatmap visualizations were generated to make the model's "reasoning" interpretable for clinicians.
  2. The Prognostic Model: This pipeline used a VGG19 architecture combined with a Cox proportional hazards regression. To address tumor heterogeneity, the model automatically identifies the 10 patches on a slide with the largest average nuclei (a hallmark of malignancy) and uses them to generate a risk score. This mimics the pathological practice of focusing on the most aggressive regions of a tumor.
Overview of the deep learning framework for RCC classification and survival prediction.

The study utilized a massive collection of 6,340 WSIs from 2,399 patients, including data from The Cancer Genome Atlas (TCGA) for primary training and internal testing, and 3,080 slides from six independent centers in China for external validation. The models were tested on both formalin-fixed paraffin-embedded (FFPE) sections and frozen sections.

Key Results

Subtype Classification

The model achieved an Area Under the Curve (AUC) of 0.979 to 0.996 in internal testing. In the multi-center external validation, it maintained high performance with AUCs ranging from 0.914 to 0.995. In a head-to-head comparison, the AI’s agreement with the ground truth (Kappa 0.839) was higher than that of two pathologists with 3 and 5 years of experience (Kappa 0.758 and 0.813, respectively).

Survival Prediction

The prognostic model achieved a c-index of 0.779, significantly outperforming the manual histologic-grade system (0.748). Patients identified as "high risk" by the model had a hazard ratio of 2.63 for poor prognosis compared to the "low risk" group, even after adjusting for age and cancer stage. By generating heatmaps, the researchers confirmed that high-risk scores were associated with known malignant features like sarcomatoid components and coagulative necrosis.

Discussion and Conclusion

This work demonstrates that deep learning can match or exceed the diagnostic accuracy of experienced pathologists while providing more objective prognostic insights. A key innovation was the "smart sampling" of regions of interest based on nuclear morphology, which proved more effective than using random patches for survival prediction.

While the results are robust, the study focused on the three most common subtypes and did not include rare variants. Future work could integrate features from immunohistochemical (IHC) stains to further improve accuracy. Nevertheless, these frameworks have significant clinical potential to assist pathologists in systematically evaluating histological information and personalizing treatment for RCC patients.

Summary updated

Abstract

<title>Abstract</title> Classifying histopathological subtypes and predicting survival of renal cell carcinoma (RCC) patients are critical steps towards treatment. In this work, we first proposed a deep learning method involving patch-based segmentation, intelligent feature extraction and heatmap visualization for classifying RCC into clear cell RCC, papillary RCC, chromophobe RCC, and adjacent benign tissue. This algorithm was trained and validated using 2,374,446 patches, 6,340 whole-slide images, 2,399 patients from The Cancer Genome Atlas and 6 independent centers. The classifiers provided areas under the curves of 0.979 to 0.996 in the internal phase, and 0.914 to 0.995 in the 6-center external phase. Furthermore, a modified deep learning approach comprising automated detection of regions of interest, patch-level learning, and morphological features-based risk scoring was developed for predicting survival of clear cell RCC patients. The prognostication model provided a hazard ratio for poor versus good prognosis of 2.63 [95% confidence interval (CI) 1.53–4.50, P = 4.35e-4] in the testing set, and 2.57 [95% CI 1.43–4.64, P = 1.68e-3] in an independent validation set using multivariable analyses. In conclusion, the developed histopathology image-based deep learning frameworks have the clinical potential to assist pathologists in systematically evaluating histological information of RCC patients.
BibTeX
@inproceedings{wang2021classifying,
  author = {Fei Wang and Yuanzhe Geng and Ting Wang and Ke Zhao and Bin Xu and Dachuan Zhang and Hongzhou Cai and Qinbo Yuan and Hang Lü and Yue Zhang and Li Xu and Yangyang Sun and Hang Gong and Raphael N. Alolga and Xiangshan Fan and Gaoxiang Ma and Lian‐Wen Qi},
  title = {{Classifying subtypes and predicting survival of renal cell carcinoma using histopathology image-based deep learning}},
  booktitle = {Research Square},
  year = {2021},
  doi = {10.21203/rs.3.rs-533678/v1},
}
Figure