Selecting Image Features for Biopsy Needle Detection in Ultrasound Images Using Genetic Algorithms

Locating the biopsy needle in ultrasound (US) images is a crucial task in medical image analysis. It aids clinicians in minimizing the risk of damaging surrounding tissue during a US-guided core needle biopsy and it allows to reduce its duration. Numerous studies have explored needle segmentation from US images, but most of them operate under the unrealistic assumption that the needle is always present in the image. To address this gap, we propose an approach for detecting the biopsy needle in US images. It couples classic machine learning with a genetic algorithm identifying the most relevant image features that contribute to needle localization. We thus concentrate on the most significant features and prune unnecessary extractors to enhance the efficiency of the pipeline which is of paramount importance in time-constrained clinical settings. Our experiments showed that genetically evolved feature subsets allow us to build effective needle detectors outperforming models trained over full feature sets, and they can be flexibly incorporated into cascaded multi-scale detection pipelines.