Method for fetal ultrasound image classification using pseudo-labelling with PCA-KMeans and an attention-augmented MobileNet-LSTM model

Abstract

Accurate classification of fetal ultrasound images is critical for early diagnosis, yet remains challenging due to limited labeled data and high inter-class variability. This study presents a robust deep learning framework that combines a MobileNet backbone with multi-head self-attention and LSTM layers to enhance feature learning and temporal context. To address data scarcity and imbalance, unsupervised clustering was employed using Principal Component Analysis (PCA) for dimensionality reduction and K-means (k=4) for pseudo-label generation. These pseudo-labeled clusters were then balanced using oversampling techniques. The proposed model was trained using transfer learning on the augmented dataset and achieved a test accuracy of approximately 98 % with a macro-F1 score of 0.98, indicating highly reliable classification performance.

• •
  Employed PCA (100 components) and K-means (k=4) for effective pseudo-labeling and class balancing.
• •
  Designed a hybrid deep learning architecture using MobileNet, multi-head attention, and LSTM.
• •
  Achieved ∼98 % test accuracy and 0.98 macro-F1 score, demonstrating strong model generalization.