Table 1.
Characteristics of patients with gastric cancer in training, validation and test set.
Fig 1.
Flowchart of the research process.
The study begins with the delineation of tumor regions from contrast-enhanced CT images. These images are used to train seven CNN architectures: ResNet, DenseNet, ShuffleNet, RegNet, Inception-ResNet, EfficientNet, and Xception, with each architecture initialized using pre-trained weights. The best-performing CNN model, Inception-ResNetV2, is then further enhanced by integrating various attention mechanism modules, including ECA (Efficient Channel Attention), CA (Coordinate Attention), SA (Spatial Attention), SE (Squeeze-and-Excitation), CBAM (Convolutional Block Attention Module), ACmix (Attention Convolution Mix), and KNNA (K-Nearest Neighbors Attention). The dataset is divided into training (60%), validation (20%), and testing (20%) sets. Model performance is evaluated using ROC-AUC, PR-AUC, F1-Score, sensitivity, specificity, and precision metrics. The final optimized model demonstrates superior performance in predicting peritoneal metastasis in patients with advanced gastric cancer.
Fig 2.
Architecture of inception-ResNetV2 model with integrated attention mechanisms.
The architecture of the optimal base model, Inception-ResNetV2, is illustrated. In this model, attention mechanisms were integrated following the first Inception-ResNet-A module, as well as after the Reduction-A and Reduction-B modules. The final model output is generated using a sigmoid activation function, producing the predicted results.
Table 2.
Performance comparison of various CNN models.
Fig 3.
ROC curves of various models on the test set.
(A) shows the ROC Curves of various CNN models on the test set; (B) illustrates the ROC Curves of the Inception-ResNetV2 model with different attention mechanisms. ROC, Receiver Operating Characteristic; CNN, Convolutional Neural Network; AUC refers specifically to the area under the ROC curve.
Fig 4.
Line plots of performance metrics across various models on test set.
(A) Figure A shows the performance metrics of various CNN models on the test set; (B) Figure B illustrates the performance metrics of the Inception-ResNetV2 model with different attention mechanisms. ROC-AUC, Receiver Operating Characteristic Area Under the Curve; PR-AUC, Precision-Recall Area Under the Curve; CNN, Convolutional Neural Network.
Fig 5.
ROC-AUC of various models at different dropout probabilities on test set.
(A) Figure A shows the ROC-AUC values of different convolutional neural network models at dropout probabilities of 0 and 0.3; (B) Figure B displays the ROC-AUC values of the Inception-ResNetV2 model with various attention mechanisms integrated at dropout probabilities of 0 and 0.3. ROC-AUC, Receiver Operating Characteristic Area Under the Curve.
Table 3.
Performance comparison of inception-ResNet V2 models with various attention mechanisms.
Fig 6.
Sensitivity and specificity across different probability thresholds for peritoneal metastasis prediction using the Inception-ResNetV2-SE model.
Sensitivity and specificity curves corresponding to various probability thresholds for peritoneal metastasis prediction using the SE-enhanced model. The x-axis represents the probability threshold (%) ranging from 0 to 100, while the y-axis shows the sensitivity (blue line) and specificity (orange line). An optimal threshold of 43% was identified, achieving a sensitivity of 95.5% and a specificity of 90.6%, providing the best balance for clinical decision-making.
Fig 7.
Calibration curves of the SE-enhanced inception-ResNetV2 model across training, validation, and test sets.
(A) Figure A displays the calibration curve for the test set; (B) Figure B displays the calibration curve for the training set; (C) Figure C displays the calibration curve for the validation set.
Fig 8.
Decision curve analysis curves of the SE-enhanced inception-ResNetV2 model across training, validation, and test sets.
(A) Figure A displays the DCA curve for the test set; (B) Figure B displays the DCA curve for the training set; (C) Figure C displays the DCA curve for the validation set. DCA, Decision Curve Analysis.
Fig 9.
Smooth grad visualization of model attention on tumor regions.
Figure A provides the visualization for a gastric cancer patient without peritoneal metastasis; Figure B showcases the visualization for a patient with peritoneal metastasis.