GastroNet FusionAI: Artificial Intelligence Diagnosis of Clinical Gastric Cancer Tumors

The significant advancements in natural image processing have sparked widespread interest in applying similar deep learning techniques to medical imaging. Unlike natural images, detecting details in medical images requires capturing finer details, necessitating the use of specially designed network architectures rather than standard ones developed for large-scale, downscaled natural images. Additionally, medical imaging often requires the fusion of images from multiple perspectives to form an accurate diagnosis. Therefore, we propose a novel AI algorithm called GastroNet FusionAI, an advanced algorithm specifically designed for gastric cancer diagnosis. It employs an innovative image fusion strategy that effectively integrates images from four different imaging perspectives in gastric cancer detection. This approach not only provides a more comprehensive view but also significantly enhances the accuracy of disease detection by enriching the image content. The development of GastroNet FusionAI aims to improve the early detection rate of gastric cancer, which is crucial for increasing patient survival rates. Gastric cancer is one of the most common types of cancer globally and has a high incidence in certain regions. Despite the enormous potential of AI technology in gastric cancer detection, research on its effectiveness and reliability remains relatively scarce. Existing studies are often limited by small sample sizes, homogeneous datasets, or a lack of rare gastric cancer subtypes. To overcome these challenges, we employed Real-Time Test-Time Augmentation (R-TTA) to expand the original dataset (consisting of 54,706 images) to 273,530 images. This method not only improved the model's learning efficiency but also effectively reduced the risk of overfitting. Our model was trained on a preprocessed dataset from three different hospitals and four different imaging perspectives, and its performance was evaluated using cross-validation to ensure the broad applicability of the research results. This included effective identification of rare gastric cancer subtypes and cases occurring in special anatomical locations. This comprehensive strategy not only significantly enhanced the overall performance of the algorithm but also greatly improved its accuracy, making GastroNet FusionAI's performance on random test subsets comparable to that of a panel of radiologists. Moreover, we compared this algorithm with five currently popular deep learning architectures, including DenseNet201, ResNet50, InceptionV3, MobileNetV2, and NASNetLarge. This comparison not only validated GastroNet FusionAI's superiority in handling complex image data but also demonstrated its ability to outperform traditional models across various evaluation metrics.