IKCS 2021: Use of AI Deep Learning in Histologic Classification  

(UroToday.com) The International Kidney Cancer Symposium 2021 annual hybrid meeting included a non-clear cell renal cell carcinoma (RCC) session and a presentation by Dr. Saeed Hassanpour discussing the use of artificial intelligence deep learning in histologic classification for renal cell carcinoma (RCC). Dr. Hassanpour started by noting that RCC is the most common renal cancer in adults, comprising ~90% of all kidney cancer. Further, histopathological characterization of RCC is important for the prognosis and treatment of patients. The motivation for artificial intelligence is that classification of RCC patterns on biopsy and surgical resection slides under the microscope are challenging given that it (i) requires highly specialized expertise, (ii) is time-consuming, (iii) and requires a high degree of variability among pathologists.

 Significant progress has been made in applying deep learning techniques, especially convolutional neural networks, to a wide range of computer vision tasks as well as biomedical imaging analysis applications. Work from Dr. Hassanpour group sought to build an artificial intelligence model to help pathologists accurately classify surgical resection and biopsy slides, and to visualize the identified indicative regions and features on digitized slides to ensure the explainability of the model.1 The histopathologic subtypes of RCC are broad, with the most common being clear cell RCC, papillary RCC, and chromophobe RCC; oncocytoma is the most common benign renal tumor type:

artificial intelligence classification for RCC-0.jpg 

This deep neural network model aimed to classify surgical resection slides into the above five histologies. Dr. Hassanpour and colleagues evaluated the model on independent test sets of 78 surgical resection whole slides and 79 biopsy slides from Dartmouth, and 917 surgical resection slides from The Cancer Genome Atlas (TCGA) database. Two pathologists manually annotated the surgical resection whole-slide images in the training and development sets. This included identifying the regions of interest for each subtype using bounding boxes, and a ground truth label for each slide was established based on original institutional labels and the verification of an additional pathologist. Overall, 727 regions of interest were identified and labelled as one of the five classes, with disagreements resolved by a senior pathologist. The deep learning model was as follows:

artificial intelligence classification for RCC-1.jpg 

The average area under the curve of the classifier on the internal resection slides, internal biopsy slides, and external TCGA slides was 0.98 (95% CI 0.97–1.00), 0.98 (95% CI 0.96–1.00) and 0.97 (95% CI 0.96–0.98), respectively. These curves are highlighted as follows (a: surgical resection whole-slide images from Dartmouth, b: surgical resection whole-slide images from TCGA, and c: biopsy whole-slide images from Dartmouth):

artificial intelligence classification for RCC-2.jpg 

Dr. Hassanpour noted the following technical contributions from this study:

  • This deep learning methodology achieved a high performance (average AUC of > 0.95) on all test sets
  • These results show a high generalizability of the approach across different data sources and specimen types
  • The visualization provides pathologists with insights into the major regions and features that contribute to the classification decisions of the method

There are several important potential clinical implications, highlighted as follows:

  • Automatically pre-screening slides to reduce false-negative cases
  • Highlighting regions of importance on digitized slides to accelerate diagnosis
  • Providing objective and accurate diagnosis as a second opinion on cases

Dr. Hassanpour concluded his presentation of using artificial intelligence deep learning in histologic classification of RCC with the following take-home messages:

  • Expanding the dataset to include rare subtypes and classes, such as clear-cell papillary renal cell carcinoma
  • Conduct prospective clinical trials to validate this approach in clinical settings and quantify its impact on the efficiency and accuracy of pathologists’ diagnosis of renal cancer
  • Develop new methods to learn signatures of the tumor microenvironment from pathology slides to inform clinical-decision making


Presented by: Saeed Hassanpour, PhD, Associate Professor of Biomedical Data Science, Epidemiology and Computer Science, Geisel School of Medicine at Dartmouth and Norris Cotton Cancer, Lebanon, NH

Written by: Zachary Klaassen, MD, MSc – Urologic Oncologist, Assistant Professor of Urology, Georgia Cancer Center, Augusta University/Medical College of Georgia, @zklaassen_md on Twitter during the International Kidney Cancer Symposium (IKCS) 2021 Annual Congress, November 5 and 6, 2021.


  1. Zhu M, Ren B, Richard R, et al. Development and evaluation of a deep neural network for histologic classification of renal cell carcinoma on biopsy and surgical resection slides. Sci Reports 2021;11:7080.


email news signup