Neurological modelling of the vision system with relevance to an application for improved detection of early breast cancer

Publication Type:
Thesis
Issue Date:
2005
Full metadata record
Files in This Item:
Filename Description Size
01Front.pdf1.87 MB
Adobe PDF
02Whole.pdf15.11 MB
Adobe PDF
Detection and recognition of early signs of breast cancer when represented by cancer-related types of microcalcifications, are key requirements in breast screening mammography. Edge detection plays a vital role not only in observing detail in lesions with microcalcifications but also for the perception of often weakly defined stellate tumours. Mammograms are read at around 60 to 80 per hour. We develop a neurological based functional model for edge responses of on-centre (Cʟ) and off-centre (Cᴅ) groups of neurons in the early vision physiology, from the retina to the striate visual cortex V1. It is shown that the responses of the resulting differential equations at the retina and at V1 are of the Lotka-Volterra (LV) type. They display two important properties: consistency with the retinotopic property of early vision physiology and susceptibility to the 1/ƒ type noise for a stochastic resonance (SR) effect on the edge responses based on Cʟ and Cᴅ interactions. Computer simulations of repeated cycles of the LV responses when 1/ƒ noise was incorporated into the interaction terms of the equations, showed an increased probability of occurrences of closely coupled small values of Cʟ and Cᴅ. This indicated an underlying stochastic resonance effect. Such occurrences suggested a method for improved edge detection for "low observables" and improved detail in microcalcification regions in mammograms. An experimental viewing system using a dynamic polaroid and eyepieces with colour filters was developed to provide 1/ƒ noise to test this hypothesis. It provided for the resulting Cʟ and Cᴅ inputs from the left and right eye of an observer to interact in the V1 laminae. Using a set of 30 microcalcification region of interest images, tests were conducted in real-time readings with five observers, each at two different times. There was double blind randomisation of the sequence of images in addition to the order of presentation of "with" and "without noise". The experience levels of the observers ranged from a low level working with technical images (not mammograms) to a high level experienced radiologist. An initial series of 300 observations using ratings of detail quality (1 to 5) showed significance for an improved effect (p<0.05). A second series of 300 observations for a more stringent test with improved symmetrical experimental design of the equipment and alternative-forced-choice for observers to reduce subjectivity, showed a "trend to significance" (p<0.1). Observers with more experience had better performance in the tests (p∼0.07). Intra-observer variability was consistently good compared with inter-observer results. A parameter in the L-V equations which is related to observer attention, coupled with a spatial search requirement, may be part of the inter-observer variability. The findings also have implications for the training of radiologists in reading mammograms in real-time screening. Recent developments in spectral properties of photonics materials may provide a simple implementation of the principles developed in the thesis.
Please use this identifier to cite or link to this item: