1441P - Clinical application of a machine learning-based pathomics signature for predicting lymph node metastasis in early gastric cancer

Background

The accurate detection of lymph node metastasis (LNM) in early gastric cancer (EGC) remains a challenge, conventional radical gastrectomy may not be warranted for EGC patients with a low LNM risk and a favorable prognosis. This study aims to construct a risk prediction model of lymph node metastasis based on clinicopathologic features, tumor budding, convolutional neural network, and machine learning in early gastric cancer.

Methods

Clinicopathological features and tumor budding (TB) were assessed using the image recognition software Qupath. There was a total of 1523471 patches, 260998 patches, and 166594 patches from three patient cohorts that were retrospectively analyzed. Convolutional neural networks (CNNs) were used for deep transfer learning to classify patches and extract features. LASSO regression and multiple machine learning algorithms were employed to combine CNN features with clinicopathological features. The risk prediction model was also verified in two independent validation cohorts.

Results

Among 200 patients, LNM incidence was 27.14% in the training cohort, 15.15% in the internal validation cohort, and 14.82% in the external validation cohort. Multivariate analysis identified pathological T stage and ITBCC-TB grade as independent LNM risk factors (p < 0.05). The Resnet152 CNN model demonstrated excellent performance with AUC values of 0.995, 0.988, and 0.965 in the training, internal validation, and external validation cohorts, respectively. The CNN-ML model using Gradient Boosting exhibited superior predictive capacity, with AUC values of 0.901 (95% CI 0.844-0.957), 0.871 (95% CI 0.726-0.999), and 0.804 (95% CI 0.555-0.999) in the training and two independent validation cohorts.

Conclusions

Pathological T stage and ITBCC-TB grade were identified as independent risk factors for LNM in EGC. The integration of clinicopathological features, image recognition, and deep transfer learning using CNNs resulted in a highly accurate predictive model for LNM in EGC. This model will provide valuable insights into the development of a more comprehensive and individualized approach to managing EGC, thereby improving the quality of care for affected patients.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.