Highlights from Medical Imaging 2007
Dr. Maria Petrou, Professor of Signal Processing at Imperial College (UK) speaks at the SPIE Medical Imaging 2007 Symposium about research currently underway by her and colleagues to model and simulate the behavior of the human retina.
Hundreds gathered on Monday evening to attend Dr. Maria Petrou's plenary presentation at SPIE Medical Imaging 2007. Dr. Petrou, currently the Professor of Signal Processing at Imperial College, gave a detailed overview of the inner workings of the human visual system, chronicled recent achievements in artificial visual system simulations, and reported on her own Research Councils UK -funded project, which is attempting to reverse engineer the entire human visual system.
She chose a system-design approach to her talk, describing first the retina and recent attempts to mimic it. The retina, Petrou explained, is quite complex in that the sensors are not directly in the path of the incident light. Instead, the human retina is composed of 5 cellular layers, including ganglion cells, bipolar cells, and horizontal cells before reaching rods and cones. This is complicated quite a bit further by the fact that the rods and cones are distributed irregularly throughout their retinal layer.
However, Dr. Petrou said, recent advances have utilized the benefits of polymers to help create retinal simulations - such advances as creation of a hybrid chip combining polymer sensors and analog circuitry, and of novel multilayer polymer deposition methods. But the irregularity of rod and cone spacing still poses a problem in terms of image processing, so a detailed physiological model of the retina has been developed to take into account data taken from irregularly spaced points.
The second part of the visual system that Dr. Petrou described was the lateral geniculate nucleus, which acts as a relay station in the brain. This passes information from the retina to the third part of the system - the visual cortex, which is responsible for human image processing.
The largest area of the visual cortex, dubbed "V1," takes the information from both eyes and interlaces the data together. So, Petrou posed, what exactly is the role of V1? How can V1 make sense of a scene, perform edge detection, and what can we learn from it to assist us in creating artificial vision?
She described recent work that has suggested that a good model for V1 is a saliency map of a given scene - essentially, V1 seems to have a means for deciding what is salient. This of course brings up the question: What is the relation between perception and vision? How does the brain decide what's important?
These questions are the very questions Petrou is attempting to answer in her current work. Coming from the signal processing perspective, she described a variety of mathematical tools she's extended from the regular to the irregular domain to allow image processing on irregular topologies. However, possibly the most interesting part of her research-and her talk-looks at the topology of ideas - the network of ideas in the brain that allows us to subconsciously identify things. A picture may be worth a thousand words, she explained, but we hardly ever find the means to verbalize a scene as well as we can simply show one. Recent studies have shown that there is much more going on subconsciously, that there is a topology of ideas in the brain that requires more exploration. Dr. Petrou discussed some of her early explorations and set the scene for all of the work yet to come.