Development of novel deep learning models for fetal movement detection in ultrasound scans

The analysis of medical images and videos with computer-based algorithms offers significant advantages in the diagnosis and treatment of diseases. Especially in recent years, advances in deep learning algorithms have led to continuous improvements in the processing of medical data in terms of speed, performance and hardware requirements. The use of deep learning algorithms to analyze highly specialized medical data has become widely used as a tool to help doctors make decisions. Tracking fetal movement and recognizing anatomical structures in ultrasound (US) videos is an important parameter in assessing infant health. Monitoring fetal movement is an important parameter in a healthy pregnancy, and reduced or absent fetal movement in the womb can be considered a sign of serious fetal risk. In this thesis, a dataset of ultrasound videos collected from Kütahya University of Health Sciences Evliya Çelebi Training and Research Hospital was created for the analysis and evaluation of fetal movements from ultrasound videos. First, a hybrid deep learning method based on YOLO and Long Short-Term Memory (LSTM) algorithms is proposed to recognize fetal anatomical structures and detect fetal movements using fetal ultrasound videos. In this phase, ultrasound videos are labelled by expert radiologists to identify fetal anatomical structures and track fetal movements. In the experimental analysis, the movements of fetal anatomical structures such as the heart, head and body are tracked and movement trajectory patterns are extracted. In the final step, the LSTM deep learning algorithm is used to detect and classify the fetal anatomical structures using the motion patterns converted to a two-dimensional (2D) plane. Automatic classification of fetal movements in ultrasound videos is crucial for assessing fetal well-being and detecting potential complications during pregnancy. In the second phase of the thesis, a new deep learning model, FetalMovNet, is developed to improve the classification of fetal movements in ultrasound video sequences by incorporating an attention mechanism and convolutional neural networks (CNN). By integrating CNNs for feature extraction and an attention mechanism to capture spatio-temporal patterns, the model significantly improves the classification performance of fetal movements. To evaluate FetalMovNet, experimental studies are conducted on a dataset of ultrasound fetal movements of seven different anatomical structures (head, body, arm, hand, heart, leg and foot). The results show that FetalMovNet outperforms advanced CNN and CNN-LSTM architectures, achieving accuracy of 0.9887, precision of 0.9871, recall of 0.9910 and F1-score of 0.9891. Ablation studies confirm the effectiveness of the attention mechanism, with FetalMovNet achieving an area under the curve (AUC) of 0.9957, compared to 0.9471 for CNN and 0.9543 for CNN-LSTM. The proposed FetalMovNet model provides a robust and clinically applicable tool for real-time fetal movement tracking, reducing the need for manual assessment and has the potential to improve prenatal care. Due to the small size of the deep learning models proposed in this thesis, the classification and detection time is very short and they are considered suitable for the evaluation of real-time ultrasound images. Another important advantage of the FetalMovNet deep learning model used in the study is its high performance even on low resolution ultrasound images. In this thesis, the classification of motion abnormalities and the detailed analysis of tissue-specific movement trajectories are considered as a promising approach for the early diagnosis of neurological disorders in the fetus. In the future, this approach may play an important role in the diagnostic process during pregnancy by being used to detect abnormal fetal movement.