Current approaches for analysis of medical data strongly rely on skilled data scientists that must be able to deliver analytical insights. However, many of these approaches present only ad-hoc solutions that have been designed to solve specific data challenges. As a result, their performance decreases dramatically if the analyzed data is highly diverse. Shifting from ad-hoc solutions towards generalized adaptive frameworks will substantially improve the analysis quality and will minimize its dependence on user’s experience. Such adaptive frameworks are of critical importance for a variety of medical applications such as treatment evaluation and precision health.
During this talk, I will present my research topics of leveraging machine learning strengths to deal with challenges that are typical for the medical domain.
I will present an adaptive segmentation technique that we developed, which can handle substantial diversity of image characteristics, and supply far more general, fast, accurate and robust segmentation solution, than any ad-hoc designed platforms that are currently available. The method is a generalization of the level set segmentation approach by developing a novel method for adaptive estimation of active contour parameters. The presented segmentation method is fully automatic once the lesion has been detected. First, the location of the level set contour relative to the lesion is estimated using a deep learning framework. Second, the output location probabilities are then used to adaptively calculate the parameters of the active contour functional during the segmentation process.
This method supplies impressive results for different imaging modalities (e.g. CT, MR, Mammography) and for different types of lesions (e.g. low contrast, heterogeneous), improving state of the art segmentation accuracy in ~27%.