Development and validation of a stacking nomogram for predicting regional lymph node metastasis status in rectal cancer via deep learning and hand-crafted radiomics

Preoperative assessment of lymph node metastasis (LNM) status is the basis of individual treatment for rectal cancer (RC). However, conventional imaging methods are not accurate enough.

We collected 282 RC patients who were divided into the training dataset (n=225) and the test dataset (n=57) with an 8:2 scale. A large number of deep learning (DL) features and hand-crafted radiomics (HCR) features of primary tumors were extracted from the arterial and venous phases of the computed tomography (CT) images. Three machine learning models, including support vector machine (SVM), k-nearest neighbor (KNN),and multi-layer perceptron (MLP) were utilized to predict LNM status in RC patients. A stacking nomogram was constructed by selecting optimal machine learning models for arterial and venous phases, respectively, combined with predictive clinical features.

The stacking nomogram performed well in predicting LNM status, with an area under the curve (AUC) of 0.914 [95% confidence interval (CI): 0.874-0.953] in the training dataset, and an AUC of 0.942 (95%CI: 0.886-0.997) in the test dataset. The AUC of the stacking nomogram were higher than those of CT_reported_N_status, ASVM, and VSVM model in the training dataset (P <0.05). However, in the test dataset, although the AUC of the stacking nomogram was higher than the VSVM, the difference was not obvious (P =0.1424).

The developed deep learning radiomics stacking nomogram showed to be effective in predicting the preoperative LNM status in RC patients.