Deep Learning Features Improve the Performance of a Radiomics Signature for Predicting KRAS Status in Patients with Colorectal Cancer

Rationale and Objectives We assess the performance of a model combining a deep convolutional neural network and a hand-crafted radiomics signature for predicting KRAS status in patients with colorectal cancer (CRC). Materials and Methods The primary cohort consisted of 279 patients with clinicopathologically confirmed CRC between April 2011 and April 2015. Portal venous phase computed tomographic images were analyzed to extract traditional hand-crafted radiomics features as well as deep learning features. A Wilcoxon rank sum test, the minimum redundancy maximum relevance algorithm, and multivariable logistic regression analysis were used to select features and build a radiomics signature. A combined model was then developed using multivariable logistic regression analysis. An independent validation cohort of 119 patients from May 2015 to April 2016 was used to confirm the combined model's predictive performance. Results The C-index of hand-crafted radiomics signature's discriminative ability was 0.719 (95% confidence interval, CI: 0.658–0.776) for the primary cohort and 0.720 (95% CI: 0.625–0.813) for the validation cohort. The C-index of the deep radiomics signature's discriminative ability was 0.754 (95% CI: 0.696–0.813) for the primary cohort and 0.786 (95% CI: 0.702–0.863) for the validation cohort. The combined model, which merged the hand-crafted radiomics features and deep radiomics features, achieve a C-index of 0.815 (95% CI: 0.766–0.868) for the primary cohort and 0.832 (95% CI: 0.762–0.905) for the validation cohort. Conclusion This study presents a model that incorporates the hand-crafted and deep radiomics signature, which can be used for individualized preoperative prediction of KRAS mutations in patients with CRC. We assess the performance of a model combining a deep convolutional neural network and a hand-crafted radiomics signature for predicting KRAS status in patients with colorectal cancer (CRC). The primary cohort consisted of 279 patients with clinicopathologically confirmed CRC between April 2011 and April 2015. Portal venous phase computed tomographic images were analyzed to extract traditional hand-crafted radiomics features as well as deep learning features. A Wilcoxon rank sum test, the minimum redundancy maximum relevance algorithm, and multivariable logistic regression analysis were used to select features and build a radiomics signature. A combined model was then developed using multivariable logistic regression analysis. An independent validation cohort of 119 patients from May 2015 to April 2016 was used to confirm the combined model's predictive performance. The C-index of hand-crafted radiomics signature's discriminative ability was 0.719 (95% confidence interval, CI: 0.658–0.776) for the primary cohort and 0.720 (95% CI: 0.625–0.813) for the validation cohort. The C-index of the deep radiomics signature's discriminative ability was 0.754 (95% CI: 0.696–0.813) for the primary cohort and 0.786 (95% CI: 0.702–0.863) for the validation cohort. The combined model, which merged the hand-crafted radiomics features and deep radiomics features, achieve a C-index of 0.815 (95% CI: 0.766–0.868) for the primary cohort and 0.832 (95% CI: 0.762–0.905) for the validation cohort. This study presents a model that incorporates the hand-crafted and deep radiomics signature, which can be used for individualized preoperative prediction of KRAS mutations in patients with CRC.