作者
Faris Kateb,Rania M. Alhazmi,Sami Saeed Binyamin,Mona Y. Alqassim,Khalid Hamed S. Allehaibi,Mahmoud Ragab
摘要
Abstract The precise classification of fetal ultrasound (US) anatomical planes is a crucial for dependable prenatal evaluation and decision-making in medical. Many metrics attained from maternal-fetal images are normally utilized for monitoring fetal development. The most widely employed biomarkers in medical training to screen fetal anomalies are fetal biometrics, prediction of fetal mass, etc. Moreover, US study procedures are time-intensive, inefficient, and prone to errors. Thus, an automated method executing this operation would improve cost efficiency and may decrease mistakes. Artificial Intelligence (AI) has experienced significant growth over the past decade, especially with the advent of deep learning (DL) and its extraordinary achievement in image classification tasks through convolutional neural networks (CNNs). In recent times, CNNs have exposed their advantage in a broad range of clinical usages, like radiology, dermatology, and/or to recognize lesions and organs from clinical images. This study proposes the Attention-Guided Graph Convolutional Network Framework for Multi-Modal Fetal Organ Classification (AGGCN-MMFOC). The proposed AGGCN-MMFOC model aims to effectively classify fetal organs captured from high-resolution ultrasound images, while providing interpretable predictions for clinically relevant anatomical planes. For accomplishing that, initially image preprocessing is performed, including image resizing and normalization, noise removal with modified anisotropic diffusion, and contrast enhancement using CLAHE. Subsequently, a learnable weighted fusion strategy is designed to integrate features from three models: EfficientNet‐B7, Twins-SVT, and Inception-V3, enabling the capture of discriminative and significant features for the classification task. The fused features are then passed to the attention-graph convolutional network for the accurate classification of six anatomical plane classes: Maternal Cervix, Fetal Abdomen, Fetal Brain, Fetal Femur, Fetal Thorax, and Other. To improve the performance of the AGGCN-MMFOC model, the parameter optimization process takes place using the Nadam optimizer for the classification model. Finally, Score-CAM is used to interpret the decision-making process of the DL approach leveraged for organ classification from ultrasound images. The experimental outcome examination of the AGGCN-MMFOC method was evaluated using various benchmark datasets. The simulation results described the improved results of the AGGCN-MMFOC system in terms of multiple metrics.