Pancreatic cancer, characterized by its poor prognosis, remains one of the leading causes of cancer deaths in the United States. Early diagnosis of pancreatic tumors is crucial for successful treatment, but their detection is challenging due to their hidden location and size. To address this, researchers have developed an automatic neural algorithm for pancreatic tumor segmentation based on deep learning. Previous efforts focused on conventional methods, but these new techniques aim to improve accuracy and reduce the burden on radiologists by automating the tumor segmentation process using advanced neural networks.
While there have been successful applications of deep learning algorithms in other domains, such as lung, liver, and heart segmentation, research on pancreatic tumor segmentation has been limited. Traditional methods have faced challenges in accurately detecting pancreatic tumors due to their low contrast and small size on medical images. The proposed algorithm addresses these challenges by introducing a two-stage neural network that leverages multi-scale U-Net and non-local localization and focusing modules to enhance the segmentation accuracy of pancreatic tumors. Additionally, the algorithm incorporates a Loss function based on the shared boundary between classes to improve the contributions of small-scale tumors to the training process.
Through extensive experiments on pancreatic tumor datasets and validation with state-of-the-art methods, the algorithm demonstrates superior performance in pancreatic tumor segmentation tasks. It outperforms other models in terms of segmentation evaluation indicators, such as the Dice coefficient, sensitivity, and specificity. The algorithm’s two-stage design, combined with advanced neural network architectures and loss functions, leads to more accurate and precise pancreatic tumor segmentation results, showing promising potential for clinical applications in cancer diagnosis and treatment.