Development of an interpretable CT radiomics model for the diagnosis of clear cell renal cell carcinoma in small solid renal masses
Development of an interpretable CT radiomics model for the diagnosis of clear cell renal cell carcinoma in small solid renal masses
Taek Min Kim, Hyungwoo Ahn, Jeong Yeon Cho, Sang Youn Kim
To develop an explainable CT radiomics-based model for diagnosing clear cell renal cell carcinoma (ccRCC) among small (≤ 4 cm) solid renal masses.
This retrospective study included 159 patients (50 women; median age 58 years [IQR 50-65 years]) with pathologically confirmed renal solid masses (≤ 4 cm). Two radiologists independently evaluated mass-to-cortex ratio and heterogeneity score (on a 5-point Likert scale) in corticomedullary phases, and assessed a five-tiered CT score for diagnosing ccRCC. Interpretable radiomics model was constructed using the CT radiomics features which were associated with mass-to-cortex ratio (first-order statistics of the mass and renal cortex) and heterogeneity score (first-order statistics and texture features of the mass). Diagnostic performance of diagnosing ccRCC were compared between five-tiered CT score and radiomics model.
The masses comprised 52.8% of ccRCC (84/159) and 47.2% (75/159) of other histologic diagnoses. The mass-to-cortex ratio and heterogeneity score were significantly higher in ccRCC than in other diagnoses (0.87 ± 0.18 vs. 0.58 ± 0.21 and 4.1 ± 0.9 vs. 2.5 ± 1.1, respectively, P<0.001 for both). The weighted kappa value of heterogeneity score between two readers was 0.64 (95% CI, 0.55-0.73). The intraclass correlation coefficient (ICC) of mass-to-cortex ratio between two readers was 0.87 (95% CI, 0.82-0.91). CT score ≥ 4 achieved an AUC, sensitivity, specificity, positive predictive value, and negative predictive value for identifying ccRCC of 0.85, 72.6%, 80%, 80.2%, and 97.5% in reader 1, and 0.83, 66.7%, 82.7%, 81.2%, and 95% in reader 2, respectively. The ICCs of mass-to-cortex ratio and heterogeneity score were 0.93 (95% CI, 0.90-0.95) and 0.82 (95% CI 0.76-0.86) between reader 1 and radiomics model, and 0.81 (95% CI, 0.74-0.86) and 0.81 (95% CI 0.74-0.86) between reader 2 and radiomics model, respectively. Diagnostic performance of radiomics model for identifying ccRCC obtained AUC of 0.91, which was superior to those of CT scores of two readers (difference between areas 0.06 and 0.08, P=0.02 and 0.01, respectively).
Distribution of ccRCC and other histolologic diagnoses using CT clear cell likelihood score in reader 1
Distribution of ccRCC and other histolologic diagnoses using CT clear cell likelihood score in reader 2
Diagnostic performance of diagnosing ccRCC using radiomics model and five-tiered CT socre in two readers
The CT-based radiomics algorithm, which was constructed using the features correlated with two key parameters, showed good performance of diagnosing ccRCC in small renal masses.
Representative image of ccRCC. Two readers assessed clear cell likelihood score of 5, and radiomics model predicted ccRCC probability of 99%.
Representative image of chromophobe RCC. Two readers assessed clear cell likelihood score of 2, and radiomics model predicted ccRCC probability of 5.6%.
