Objective:

To increase the prediction accuracy of pathological nodal stage after neoadjuvant chemoradiotherapyby using the interaction between tumor features and lymph node features from pretreatment magnetic resonance imaging (MRI) data.

Methods:

The MRI data and pathological N-stage of 229 patients with locally advanced rectal cancer were retrospectively collected. And 3 radiologists delineated region of interest on tumor region and lymph node regions. 41 features were extracted from the primary tumor and the largest lymph node respectively. 41 interaction terms were calculated. Receiver operating characteristic(ROC) curve was generated by comparing each of the 123 features with the pathological N-stage. Area under curve (AUC) was calculated.

Results:

After ranking the 123 features with a descending order of AUC value

21 out of the first 41 features belong to the interaction terms. The contract feature in gray level cooccurrence produces AUC=0.727 6 at the interaction term

larger than using the corresponding tumor feature (AUC=0.555 3) or lymph node feature (AUC=0.713 9) alone. The signal sum in first-order features produces AUC=0.713 5 at the interaction term

larger than using the corresponding tumor feature (AUC=0.567 6) or lymph node feature (AUC=0.690 8) alone.

Conclusion:

The interaction between tumor features and lymph node features from pretreatment MRI data performed better than each of them for the prediction of pathological N-stage after neoadjuvant chemoradiotherapy.