The value of machine learning based on CT radiomics in the preoperative identification of peripheral nerve invasion in colorectal cancer: a two-center study

Abstract

Background: We aimed to explore the application value of various machine learning (ML) algorithms based on multicenter CT radiomics in identifying peripheral nerve invasion (PNI) of colorectal cancer (CRC).

Methods: A total of 268 patients with colorectal cancer who underwent CT examination in two hospitals from January 2016 to December 2022 were considered. Imaging and clinicopathological data were collected through the Picture Archiving and Communication System (PACS). The Feature Explorer software (FAE) was used to identify the peripheral nerve invasion of colorectal patients in center 1, and the best feature selection and classification channels were selected. Finally, the best feature selection and classifier pipeline were verified in center 2.

Results: The six-feature models using RFE feature selection and GP classifier had the highest AUC values, which were 0.610, 0.699, and 0.640, respectively. FAE generated a more concise model based on one feature (wavelet-HLL-glszm-LargeAreaHighGrayLevelEmphasis) and achieved AUC values of 0.614 and 0.663 on the validation and test sets, respectively, using the "one standard error" rule. Using ANOVA feature selection, the GP classifier had the best AUC value in a one-feature model, with AUC values of 0.611, 0.663, and 0.643 on the validation, internal test, and external test sets, respectively. Similarly, when using the "one standard error" rule, the model based on one feature (wave-let-HLL-glszm-LargeAreaHighGrayLevelEmphasis) achieved AUC values of 0.614 and 0.663 on the validation and test sets, respectively.

Conclusions: Combining artificial intelligence and radiomics features is a promising approach for identifying peripheral nerve invasion in colorectal cancer. This innovative technique holds significant potential for clinical medicine, offering broader application prospects in the field.

Critical relevance statement: The multi-channel ML method based on CT radiomics has a simple operation process and can be used to assist in the clinical screening of patients with CRC accompanied by PNI.

Key points: • Multi-channel ML in the identification of peripheral nerve invasion in CRC. • Multi-channel ML method based on CT-radiomics can detect the PNI of CRC. • Early preoperative identification of PNI in CRC is helpful to improve the formulation of treatment strategies and the prognosis of patients.

Keywords: Colorectal cancer; Computed tomography; Machine learning; Perineural invasion; Radiomics.

Fig. 1

The inclusion and exclusion criteria of patients

Fig. 2

The schematic diagram of the entire radiomics and machine learning pipeline for three cohorts of patients

Fig. 3

For the model performance generated by RFE, CV Train represents the average result of the K-1 folder training data set in K-folder cross-validation. CV validation represents the average result of the 1-folder dataset of the training set in K-folder cross-validation, Train represents the result by all training sets, and Test represents the result of the test set. a Receiver operating characteristic (ROC) curves of the model with different data sets. b Feature Explorer (FAE) software proposes candidate feature models according to the “one standard error” rule. c Features selected in diagnosis

Fig. 4

Model performance generated by RFE. a Receiver operating characteristic (ROC) curves of the model with different data sets. b Feature Explorer (FAE) software proposes candidate feature models according to the “one standard error” rule. c Features selected in diagnosis

Fig. 5

Model performance generated by ANOVA. a Receiver operating characteristic (ROC) curves of the model with different data sets. b Feature Explorer (FAE) software proposes candidate feature models according to the “one standard error” rule. c Features selected in diagnosis

Fig. 6

RFE-generated model performance, receiver operating characteristic (ROC) curves of the model under different data sets. a The ROC curve was evaluated in center 2 using RFE feature selection and the GP classifier. b The ROC curve was verified in center 2 using ANOVA feature selection and the GP classifier