Abstract
Objective The aim of the study was to assess and compare dual-energy CT (DECT) based quantitative
parameters to differentiate between clear cell renal cell carcinoma (ccRCC) and non-ccRCC.
Materials and Methods This was a retrospective study including RCC patients who underwent DECT prior to
surgery between January 2017 and December 2022. Two DECT parameters—iodine concentration
(IC) and iodine ratio (IR)—were measured by two independent readers who manually drew
circular regions of interest on the most enhancing part of the tumor. Inter-reader
agreement was calculated using the intraclass correlation coefficient. Machine learning
(ML) models trained to classify the histologic subtype as ccRCC and non-ccRCC, and
grade of ccRCC as low or high, were evaluated for their accuracy.
Results A total of 112 patients (mean age: 65 years; male:female: 61:51), with 87 ccRCCs
and 25 non-ccRCCs, were included. There was good inter-reader agreement for both IC
and IR with a Pearson coefficient of 0.89. The individual DECT parameters had an accuracy
of 77.7% (IC) and 77.5% (IR) for distinguishing ccRCC and non-ccRCC. Random Forest
classifier and AdaBoost were the best ML models with an accuracy of 89.2% each. When
ML algorithms were combined, the performance was improved, with AdaBoost performing
the best with an accuracy of 100%. To distinguish low- and high-grade ccRCCs, IC and
IR had an accuracy of 77.9 and 77.6%, respectively, while the ML models all did equally
well with an accuracy of 77.6%. Combining ML algorithms again led to improved performance,
with AdaBoost being the best overall ML model.
Conclusion DECT-based quantitative imaging biomarkers have moderate diagnostic accuracy, which
can be greatly improved using ML to differentiate between ccRCC and non-ccRCC and
predict the grade of ccRCC.
Keywords
clear cell renal cell carcinoma - RCC subtyping - machine learning
