Scholarlink

Objective <p>Postoperative metachronous liver metastasis (MLM) in colorectal cancer (CRC) patients is often difficult to predict using conventional clinical and radiological methods, which may result in delayed diagnosis and treatment. We aimed to develop and validate an artificial intelligence integrated model to improve MLM prediction after CRC surgery.</p> Methods <p>A retrospective study was analyzed (<i>n</i> = 522) CRC patients underwent for radical surgery between 2014 and 2019. Categorized into MLM (<i>n</i> = 106) and non-MLM (<i>n</i> = 416) groups based on the presence of postoperative liver metastasis within 5 years. The dataset was split 8:2 for training and validation, utilizing 5-fold cross-validation. Data included demographic factors, tumor characteristics, laboratory results, and CT arterial-phase images. Feature selection employed Random Forest Boruta and Lasso Regression with 5-fold cross-validation to identify key predictors. A Multimodal model integrated numerical and image data (CMLM) was developed, integrating value-based features through a Self-Attention Dense ResNet (SAD) module and image-based features through a Convolution Vision Transformer (CVT) module. Features extracted from SAD and CVT were fused by feature splicing, and MLM was predicted by full connection layer. Model performance was assessed by ROC curves, calibration curves, decision curves, and survival analyses. Interpretability was enhanced through Shapley values for numerical data and Grad-CAM for imaging data.</p> Results <p>The fusion model CMLM predicts the accuracy of MLM at 0.88 (0.84–0.91), with a recall rate of 0.80 (0.75–0.84), an F1 score of 0.78 (0.75–0.81), and an AUC value of 0.85 (0.84–0.86), which is higher than the accuracy, recall rate, F1 score, and AUC values of the single-modality models SAD the performance metrics for the model are as follows: an accuracy rate of 0.71 (0.56–0.86), a recall rate of 0.60 (0.53–0.68), an F1 score of 0.55 (0.45–0.65), and the AUC value of 0.69 (0.68–0.69). In comparison, the CVT model exhibits an accuracy rate of 0.66 (0.50–0.83), recall rate of 0.65 (0.63–0.67), F1 score of 0.57 (0.49–0.65), and the AUC value of 0.72 (0.70–0.73), with a statistically significant difference (<i>p</i> &lt; 0.001). The top five key predictive factors for SAD included ANC, ALB, ALC, pT stage, and PNI. Grad-CAM highlighted key regions for predicting MLM in imaging information of the preoperative primary lesions, showing similar results in independent external data validation, exhibits robust generalization.</p> Conclusion <p>The model that integrates clinical modalities and primary tumor lesion modalities can significantly improve the prediction of MLM after CRC surgery, providing a valuable tool for early detection, diagnosis, and treatment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multimodal prediction of metachronous liver metastasis in stage I-III colorectal cancer patients: multicenter cohort study employing machine learning

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Lei Liang,
  • Yahan Zhang,
  • Liuyang Yang,
  • Junnan Li,
  • Tawfik Ali Hamood Alburiahi,
  • Wanrong Lin,
  • Yanhong Yang,
  • Ruize Zhou,
  • Zhenya Yang,
  • Xihong Liu,
  • Zhengqi Wen,
  • Ning Xu,
  • Liyu Shan,
  • Jun Yang

摘要

Objective

Postoperative metachronous liver metastasis (MLM) in colorectal cancer (CRC) patients is often difficult to predict using conventional clinical and radiological methods, which may result in delayed diagnosis and treatment. We aimed to develop and validate an artificial intelligence integrated model to improve MLM prediction after CRC surgery.

Methods

A retrospective study was analyzed (n = 522) CRC patients underwent for radical surgery between 2014 and 2019. Categorized into MLM (n = 106) and non-MLM (n = 416) groups based on the presence of postoperative liver metastasis within 5 years. The dataset was split 8:2 for training and validation, utilizing 5-fold cross-validation. Data included demographic factors, tumor characteristics, laboratory results, and CT arterial-phase images. Feature selection employed Random Forest Boruta and Lasso Regression with 5-fold cross-validation to identify key predictors. A Multimodal model integrated numerical and image data (CMLM) was developed, integrating value-based features through a Self-Attention Dense ResNet (SAD) module and image-based features through a Convolution Vision Transformer (CVT) module. Features extracted from SAD and CVT were fused by feature splicing, and MLM was predicted by full connection layer. Model performance was assessed by ROC curves, calibration curves, decision curves, and survival analyses. Interpretability was enhanced through Shapley values for numerical data and Grad-CAM for imaging data.

Results

The fusion model CMLM predicts the accuracy of MLM at 0.88 (0.84–0.91), with a recall rate of 0.80 (0.75–0.84), an F1 score of 0.78 (0.75–0.81), and an AUC value of 0.85 (0.84–0.86), which is higher than the accuracy, recall rate, F1 score, and AUC values of the single-modality models SAD the performance metrics for the model are as follows: an accuracy rate of 0.71 (0.56–0.86), a recall rate of 0.60 (0.53–0.68), an F1 score of 0.55 (0.45–0.65), and the AUC value of 0.69 (0.68–0.69). In comparison, the CVT model exhibits an accuracy rate of 0.66 (0.50–0.83), recall rate of 0.65 (0.63–0.67), F1 score of 0.57 (0.49–0.65), and the AUC value of 0.72 (0.70–0.73), with a statistically significant difference (p < 0.001). The top five key predictive factors for SAD included ANC, ALB, ALC, pT stage, and PNI. Grad-CAM highlighted key regions for predicting MLM in imaging information of the preoperative primary lesions, showing similar results in independent external data validation, exhibits robust generalization.

Conclusion

The model that integrates clinical modalities and primary tumor lesion modalities can significantly improve the prediction of MLM after CRC surgery, providing a valuable tool for early detection, diagnosis, and treatment.