The aim of this study was to explore the feasibility of a high-resolution T2-weighted imaging (HR-T2WI)-based radiomics prediction model for diagnosing metastatic lymph nodes (LNs) within the mesorectum in rectal cancer.
A total of 604 LNs (306 metastatic and 298 non-metastatic) from 166 patients were obtained. All patients underwent HR-T2WI examination and total mesorectal excision (TME) surgery. Four kinds of segmentation methods were used to select region of interest (ROI), including method 1 along the border of LNs; method 2 along the expanded border of LNs with an additional 2–3 mm; method 3 covering the border of LNs only; and method 4, a circle region only within LNs. A total of 1,409 features were extracted for each method. Variance threshold method, Select K Best, and Lasso algorithm were used to reduce the dimension. All LNs were divided into training and test sets. Fivefold cross-validation was used to build the logistic model, which was evaluated by the receiver operating characteristic (ROC) with four indicators, including area under the curve (AUC), accuracy (ACC), sensitivity (SE), and specificity (SP). Three radiologists with different working experience in diagnosing rectal diseases assessed LN metastasis respectively. The diagnostic efficiencies with each of four segmentation methods and three radiologists were compared to each other.
For the test set, the AUCs of four segmentation methods were 0.820, 0.799, 0.764, and 0.741; the ACCs were 0.725, 0.704, 0.709, and 0.670; the SEs were 0.756, 0.634, 0.700, and 0.589; and the SPs were 0.696, 0.772, 0.717, and 0.750, respectively. There was no statistically significant difference in AUC between the four methods ( p > 0.05). Method 1 had the highest values of AUC, ACC, and SE. For three radiologists, the overall diagnostic efficiency was moderate. The corresponding AUCs were 0.604, 0.634, and 0.671; the ACCs were 0.601, 0.632, and 0.667; the SEs were 0.366, 0.552, and 0.392; and the SPs were 0.842, 0.715, and 0.950, respectively.