[Contrast-enhanced CT-based habitat radiomics for analyzing the predictive capability for oral squamous cell carcinoma].

OBJECTIVES

By comparing deep learning and habitat analysis models based on contrast-enhanced CT (CECT), this study explores a novel approach to predict cervical lymph node metastasis and pathological subtypes in oral squamous cell cancer (OSCC).

METHODS

CECT images from patients diagnosed with OSCC by paraffin pathology were retrospectively collected. A total of 107 patients underwent primary lesion resection and cervical lymph node dissection. Region-of-interest images under CECT were divided into three regions using K-means clustering, and feature selection was performed through a fully connected neural network to construct a habitat analysis model. A clinical model was constructed using nine clinical features, including age, gender, and tumor location. With pathological subtypes and lymph node metastasis (LNM) as study endpoints, the predictive capabilities of the clinical model, deep learning model, habitat analysis model, and combined clinical + habitat model were compared using confusion matrices and receiver operating characteristic (ROC) curve.

RESULTS

The habitat-clinical combined model exhibits superior predictive performance: in the prediction of lymph node metastasis, the area under the receiver operating characteristic curve (AUC) reaches 0.97. In the prediction of pathological typing, the AUC values are 1.00 for well-differentiated type, 0.94 for moderately differentiated type, and 1.00 for poorly differentiated type. This combined model outperforms the standalone clinical and habitat models in predicting pathological typing and lymph node metastasis.

CONCLUSIONS

The habitat-clinical integrated model exhibits superior predictive efficacy in evaluating LNM and pathological classification in oral carcinoma.