Analyzing the evaluation metrics of detecting gastrointestinal tumor using segmentation techniques in endoscopic images

Convolutional Neural Networks (CNNs) are gaining popularity for analyzing endoscopic images due to their many benefits. Since certain gastric polyps can lead to stomach cancer, it's crucial to detect and remove them accurately and promptly. CNN-based semantic segmentation helps by precisely outlining polyp areas, aiding endoscopists in identifying and treating them effectively. Despite the potential benefits, there is a scarcity of studies employing CNN for automated gastric polyp identification, particularly in the realm of semantic segmentation. Thus, we present groundbreaking research focused on segmenting gastric polyps in endoscopic images using CNNs. Various traditional semantic segmentation models, such as U-Net, DeepNet, SegNet, FuNet, and CustomNet (referred to as GISTNet), employing encoders like U-Net, ResNet50, MobileNetV2, or EfficientNet-B1, were constructed and scrutinized using a comprehensive dataset. Given the complexity of the
problem and the multitude of criteria, selecting the most suitable CNN model poses a challenge. To address this, we propose an integrated evaluation approach that combines subjective considerations with objective data to
identify the optimal CNN model. Our proposed network, CustomNet (GIST Net), employing ResNet as the encoder, emerged as the top performer according to our integrated evaluation method and was selected to construct the automated polyp segmentation system. This investigation underscores the clinical significance of semantic segmentation models in gastric polyp diagnosis and highlights the efficacy of the integrated evaluation approach in impartially selecting suitable models. Additionally, our research has the capacity to progress the identification techniques of gastric cancer, and the
proposed evaluation methodology has implications for selecting diagnostic techniques based on mathematical models.