Next-gen ultrasound imaging by closing the perception-action loop

Descripción del proyecto

Mejorar la calidad de las ecográficas: empezar por el final

Mejorar la calidad de las ecografías de bajo coste facilitaría el acceso al diagnóstico. El equipo del proyecto US-ACT, financiado por el Consejo Europeo de Investigación, considera la ecografía como un dispositivo de percepción y acción, cuya acción es la adquisición de información sobre el cuerpo humano y cuya percepción es la reconstrucción de esa anatomía. En el proyecto se pretende mejorar la calidad de la imagen eliminando el bucle entre percepción y acción mediante el uso de la percepción para impulsar la adquisición. Gracias a las matemáticas y al concepto de dimensionalidad múltiple, se obtendrán modelos de percepción sólidos basados en avanzadas redes generativas profundas. A partir de la anatomía pura descrita por el modelo generativo, la adquisición identificará el punto en su baja dimensionalidad múltiple: los datos reducidos que conservan la precisión requerida para la tarea.

Objetivo

Ultrasound (US) can revolutionize and democratize medical imaging if it offers: (1) access for everyone, and (2) excellent Image Quality (IQ). MRI offers (2) but is expensive and will thus not likely be able to provide (1). Low-cost US hardware technology will enable (1) in the future but is not expected to yield the needed breakthrough for (2). Consequently, any paradigm-shifting advance in signal processing technology that achieves US with excellent IQ will have a huge impact.

I propose a conceptually new and highly-unconventional approach that I believe can lead to a new generation of US technologies with excellent IQ. I will formally describe US systems as intelligent autonomous agents that perform actions and perception using probabilistic inference: the action is the acquisition, probing the world, and the perception is the reconstruction that infers what anatomy most likely generated the acquired US data. I conclude that current US systems are in essence flawed agents since (1) actions are not driven by perception, i.e. the perception-action loop is broken, and (2) their generative perception models are naive. My proposal will address this by closing the perception-action loop and offering strong perception models based on advanced deep generative networks. This breaks a fundamental tenet in US imaging, where I put forth the important concept that the acquisition and perception should work together to identify the point on the low-dimensional manifold of pure anatomy (described by the generative model) that is being imaged.

My intelligent US agents will pursue excellent IQ under the heading of a single probabilistic principle: minimization of ``surprise’’ under the agent’s own prior belief (the generative model) that such high-quality images can indeed be achieved. With this, we open a new frontier within active imaging (in US and beyond) where data acquisition and information processing are treated jointly based on expressive generative density functions.