Artificial intelligence-assisted noninvasive preoperative prediction of lymph nodes metastasis in cervical cancer through a clinical-imaging feature combined strategy.

BACKGROUND

Lymph node metastasis (LNM) of patients with cervical cancer (CC) is correlated with noticeably reduced five-year survival rate. but the role of conventional detection is limited for preoperative diagnosis of LNM. Therefore, we intended to develop a predictive model for LNM by integrating medical images, clinical data along with artificial intelligence-assisted method.

METHODS

CC patients who underwent radical hysterectomy combined with pelvic lymphadenectomy between January 2013 and October 2024 were retrospectively enrolled in this study. For computed tomography (CT) and ultrasound (US) images, a pre-trained ResNet-18 model on large-scale samples was used to extract representative features, fine-tuned with random cropping data augmentation. For clinical indicators, after normalizing to the range [0,1], a multilayer perceptron block was applied to extract representative features. Then, contrastive learning and feature fusion methods were utilized to integrate similar messages. Finally, a multi-modal contrastive learning framework was developed by consolidating above two parts. The framework was estimated by accuracy, sensitivity, specificity and the area under the receiver operating characteristic curve (AUC).

RESULTS

This work consisted of 127 CT images of patients with pathologically diagnosed cervical malignancies. After integrating clinical-imaging feature and artificial intelligence-assisted algorithm, the finally developed LNM predicting model achieved a high accuracy of 92.31% with an AUC of 0.88. Additionally, the model also displayed strong sensitivity (80.0%) and specificity (95.45%) in CC cohorts.

CONCLUSION

This study presented an efficient noninvasive and highly accurate diagnostic tool for LNM, which may significantly enhance surgical decision-making for lymph node dissection in CC patients with LNM.