FETR: A Weakly Self-Supervised Approach for Fetal Ultrasound Anatomical Detection

Weakly supervised object detection (WSOD) is a cutting-edge research field within computer vision that aims to detect objects in images with minimal or incomplete annotations. In this regard, we propose a novel WSOD architecture optimized for fetal ultrasound imaging. The model is designed to leverage the inherent capabilities of the convolutional neural network and the transformers to localize and classify fetal anatomical structures within ultrasound images without requiring extensive annotated datasets. Employing a class-agnostic Fetal Transformer (FETR) for generating high-quality object proposals, our approach integrates a Multiple Instance Learning (MIL) framework to enhance detection sensitivity. We conduct thorough experiments on the FPUS23 dataset, incorporating strategic data augmentation techniques to ensure model robustness while maintaining the diagnostic integrity of the images. The efficacy of our method is demonstrated through extensive evaluations, where it achieves superior performance against state-of-the-art models, not only on the FPUS23 dataset but also on the Fetal Plane DB dataset, showcasing its adaptability to various imaging conditions. The results from our ablation studies further validate the significance of each architectural component, with qualitative results emphasizing the model's precision and reliability. Our work sets the stage for advanced prenatal diagnostics, promising to elevate the standards of fetal health monitoring and care.