Application of U-Net Network Algorithm in Electronic Information Field

Abstract

This rapidly evolving landscape, which includes the field of medical diagnostics, has integrated with the electronic data (E-Data) field to provide precise and efficient treatment for complex medical conditions. The research field has further catapulted its reach to include various data types, including image, video, medical expert diagnostic type, and sensor input, out of which the image-based diagnostic model has excellent research potential. Convolutional Neural Network (CNN) based models have evolved into better Deep Learning (DL) models for handling complex intricacies featured in the input image. U-Net is a prominent CNN model developed to handle the features of image data. The U-Net excels in capturing detailed features through its encoder-decoder structure and skip connections, but its uniform weighting across different network layers may not adequately address the subtleties involved in complex medical anomaly detection. This work proposed the Attention Calibrated U-Net (ACU-Net) model that is designed to address the challenges of U-Net in detecting Fetal Cardiac Rhabdomyoma (FCR) from echocardiographic (ECG) images. FCR is a prevalent benign cardiac tumor in fetuses that poses significant diagnostic challenges due to its variable manifestations and the intricate nature of fetal cardiac anatomy. The proposed model enhances the U-Net architecture with attention mechanisms and employs a hybrid Loss Function (LF) that combines Cross-Entropy Loss, Dice Loss, and an attention-driven component for effective FCR detection. The model was compared against others and demonstrated better specificity, accuracy, precision, recall, and F1-score performance across various ECG views (LVOT, RVOT, 3VT, and 4CH).