Noninvasive KRAS mutation estimation in colorectal cancer using a deep learning method based on CT imaging.

Area of Science:

• Radiology
• Oncology
• Artificial Intelligence

Background:

• Accurate Kirsten rat sarcoma viral oncogene homolog (KRAS) mutation status is crucial for tailoring colorectal cancer (CRC) treatment.
• Noninvasive prediction of KRAS status in CRC remains a significant clinical challenge.
• Deep learning (DL) methods show promise in medical image analysis for diagnosis and prediction.

Purpose of the Study:

• To investigate the predictive performance of a DL model using residual neural network (ResNet) for estimating KRAS mutation status in CRC patients.
• To assess the utility of pre-treatment contrast-enhanced CT imaging for noninvasive KRAS mutation prediction.
• To compare the predictive ability of the DL model against a radiomics model.

Main Methods:

• A dataset of 157 CRC patients was used, with 117 in the training cohort and 40 in the testing cohort.
• A ResNet model was developed using portal venous phase CT images (axial, coronal, sagittal) to predict KRAS mutations.
• Expanded regions of interest (ROI) and a radiomics model with random forest classifier (RFC) were explored for comparison.

Main Results:

• The ResNet model achieved an area under the curve (AUC) of 0.90 in the axial direction, peaking at 0.93 with an expanded ROI.
• The radiomics model achieved an AUC of 0.818 in the testing cohort.
• The ResNet DL model demonstrated superior predictive ability compared to the radiomics model.

Conclusions:

• Computerized assessment of pre-treatment CT images using a DL model can accurately predict KRAS mutations in CRC patients.
• This DL approach offers a potential noninvasive method for estimating KRAS mutation status.
• The findings support the use of DL in medical imaging for personalized cancer therapy planning.