Large Data Sets Workshop University of Nottingham 13 th April 2006 Polarized Light Imaging for Skin Cancer Diagnosis James Housley School of Electrical and Electronic Engineering University of Nottingham
Overview Skin and polarized light Data so far Analysis methods so far Ideal data What can we do with it? Conclusion
Skin is permeable to light Surface Reflections Superficial Visitation Deep Visitation Light In (Not to scale)
Skin is permeable to light Light In Light Out
Polarized Light Linear Polarization Circular Polarization
Co-polar and Cross-polar Co-polarCross-polar Co-polarReference Linear: Reference Circular:
How is polarization useful? More collisions, less polarization maintained Deeper visitation, more collisions Deeper visitation = less polarization maintained
Linearly polarized light on skin Co-polar Random (Co-polar + Cross-polar) Linearly Polarized Light In
How can we use that? Detect co-polar and cross-polar light separately Channel 1 – channel 2: surface + superficial + deep – deep = surface + superficial ChannelLight ConfigurationSkin Information 1Linear Co-polarSurface Reflections & Superficial Layers & Deep Layers 2Linear Cross-polarDeep Layers
Linear vs. circular Surface reflections are cross-polar in circular polarization compared to co-polar in linear polarization For circularly polarized light, the direction of polarization is maintained, but the direction of the light is reversed. Therefore circular polarization is ‘flipped in helicity’ by reflections Linearly polarized light stays polarized in the same plane after reflection Light Polarization
Circularly polarized light on skin Cross-polar (cf. co-polar for linear polarization) Co-polar Random (Co-polar + Cross-polar) Circularly Polarized Light In
Any better? Channel 3 – channel 2: superficial + deep – deep = superficial ChannelLight ConfigurationSkin Information 1Linear Co-polarSurface Reflections & Superficial Layers & Deep Layers 2Linear Cross-polarDeep Layers 3Circular Co-polarSuperficial Layers & Deep Layers 4Circular Cross-polarSurface Reflections & Deep Layers
A demonstration Channel 1Channel 3Channel 4Channel 2 Channel 1 – 2Channel 3 – 2
What next? Extract information from images Malignant Lesions Benign Lesions Comparing Channels Scattergraph - every point represents the intensity of a pixel in two different channels
Comparing channels
Principal components analysis Method of reducing dimensions in data Four images = four dimensions 1 st principal component is an image which contains the most possible information from all four images Represents the best possible way of reducing the four dimensional data down to one dimension
Principal components analysis
Ideal data 4 channels 4 light sources 16 images per skin sample Or, for superficial skin layers only 4 images per skin sample (1 per light source)
What can we do with this data? Principal components analysis Segmentation Neural networks
Conclusion ?
Acknowledgements Dr. Steve Morgan Dr. John Crowe Dr. Ian Stockford
Any questions? Channel 1Channel 3Channel 4Channel 2 Channel 1 – 2Channel 3 – 2