(UroToday.com) In the Top Abstracts session of the 2021 International Kidney Cancer Symposium (IKCS): North America meeting, Dr. Spencer presented her work, based on the TRACERx Renal project, to examine the use of machine learning to predict BAP1/PBRM1 using digital histology in patients with clear cell renal cell carcinoma (ccRCC).
Beginning, Dr. Spencer emphasized that ccRCC has a relatively unpredictable clinical course. Among patients with localized disease, some patients will have a non-progressive course whereas others have relatively rapid progression. Among those who present with more advanced disease, there is again a wide range of clinical phenotypes. Currently, it is very difficult to separate indolent from aggressive cancers.
She further emphasized that disease progression is driven by genetic alterations. Understanding the patterns of these genetic alterations may help us better predict patient outcomes. In particular, she noted that tumor evolution can occur by three different patterns. Each of these are associated with characteristic clinical patterns.
Among these, patients with linear evolution have the lowest intratumoral heterogeneity and the best clinical prognosis. In contrast, those with punctuated evolution have a relatively poor prognosis. BAP1 mutations are associated with this punctuated evolution.
Currently, histological analysis represents the diagnostic gold standard. However, better understanding molecular alterations from histology may allow us to improve clinical management. The substantial intratumoral heterogeneity in ccRCC precludes current clinical utility of genomic profiling. However, this heterogeneity can be captured through histology. Thus, if we could predict genetic alterations from histology may provide a cost-effective and implementable solution.
Thus, the authors examined 94 ex vivo samples of ccRCC with multi-regional sampling of each tumor based on the TRACERx Renal cohort. This cohort has a remarkable ability to capture intra-tumoral heterogeneity due to regional sampling of the tumor. Further, mirror image sampling of bisected tumors links genomic and histologic data.
Among 50 patient cases and 606 primary tumor regions, the authors used a convolutional neural network to predict, based on image processing, the presence of mutations in BAP1 and PBRM1.
This model, when examined in a hold-out testing cohort, had very good performance characteristics, with AUCs of 0.85 and 0.88, the accuracy of 0.967 and 0.926, and F-1 scores of 0.929 and 0.925, for BAP1 and PBRM1, respectively. Dr. Spencer highlighted that these data represent an exciting proof of principle that CNN assessment of histology can characterize genomic characteristics. Such an approach could be used to assess other important molecular alterations and may allow for biological insights.
Presented by: Charlotte Spencer, MBBS, MRCP