Maternal Fetal Ultrasound Planes Classification using Contrastive Language Image Pre-training Models

Ultrasound images can be used to monitor the progress and health of fetal organs. Identification of fetal planes from the ultrasound videos is cumbersome and error prone. A recent work on fetal plane classification demonstrated some encouraging results towards this. An improved algorithm using state of the art Contrastive Language Image Pre-training models (CLIP) with the help of Synthetic Minority Over Sampling techniques is proposed in this work. The selected dataset comprises 12400 images with six classes. Number of images in each class are not same and hence imbalanced metrics are used. Image embeddings/features are extracted by passing the ultrasound images through pre-trained CLIP models which are trained on millions of image-text pairs. Diverse architectures trained on different datasets are available with CLIP models. 64 CLIP models are evaluated and their ability to classify the maternal fetal ultrasound planes is presented. Out of 64 models chosen, ViT-BigG-LAION2B offered highest accuracy. As the dataset is an imbalanced one, various over samplers are used and the model is trained with two over samplers. Results are compared with a recent work on same dataset, and the superiority of the proposed method is justified.