Fetal plane classification from 2D-ultrasound images leveraging squeeze and excitation self-attention mechanism for feature recalibration in MedMamba

Abstract

A fetal ultrasound is a safe pregnancy test that provides an image of the baby’s heart, head, and spine while also analyzing various aspects of its anatomy. Maternal-fetal ultrasound imaging is critical during pregnancy, but existing approaches rely on manual interpretation, which can be time-consuming and can overlook irregularities. Thus, the exploration of fetal ultrasound imaging has resulted in the need for accurate and fast medical image classification. However, there have been some limitations with traditional methods such as Convolutional Neural Networks (CNNs) and transformer models. For example, CNNs do not work well when it comes to modeling long-range dependencies that are very important in medical image feature extraction. Also, transformers have a high quadratic complexity hence demanding too much computation despite being good at dealing with long-range interactions. Our thesis is inspired by recent advances made in state space models (SSM) and thus presents an implementation of Vision Mamba called “MedMamba” designed for classifying medical images. This is achieved through integration of SS-Conv- SSM module which combines local feature extraction capabilities brought about by convolution layers together with long range dependency modeling as exhibited by SSMs thus solving the above mentioned problems encountered during CNN usage. In other words we can say that this hybrid method guarantees strong feature extraction across different types of medical imaging modalities while improving on computational efficiency. Furthermore we have presented an enhancement of the architecture MedMamba called “MedMambaSE” by adding a Squeeze and Excitation (SE) block. This addition refines the process of recalibrating features ultimately improving the models sensitivity and accuracy in detecting abnormalities in development. The incorporation of this block boosts MedMambas adaptability and effectiveness, in handling the complexities of ultrasound images. Through experiments on a dataset of ultrasound images we have shown that MedMambaSE not only enhances classification accuracy but also establishes a new standard for automated analysis of fetal images. This study sets a milestone, in diagnostics and opens doors for advancements in AI driven medical imaging that could revolutionize prenatal care with quicker and more precise interpretations.