Texture and colour for automatic image-based skin lesion analysis
The research presented here considers automatic diagnosis support for skin cancer. The role of computer-based diagnosis, and its value within a primary care situation are examined resulting in synthesis of aims, requirements and properties for an effective system -a system based on digital optical images captured and processed using low-cost commercial computer technology. The issues involved in acquisition of lesion boundaries are discussed. The value of accurate and robust boundaries, in terms of both directly obtainable diagnostic features and in enabling lesion property evaluation, is identified. Previous research has proposed the edge focusing process. This work has addressed the improvement, in terms of potential for future development, evaluation and reuse, of this process through porting it to a highly modular form in the Khoros environment. The role of colour analysis and its value in terms of provision of diagnostically useful features is investigated, and the central importance of segmentation is identified. The fundamental properties of effective segmentation of lesion image colours are identified as a need to reflect human perception of colour similarity and a basis on local regions. A new region-based segmentation technique using data transformed to a perception-uniform colour-space is presented and shown to yield promising results. Finally the use of texture information is discussed. The nature and properties of the large-scale texture of skin patterning and its disruption are investigated and an abstracted representation proposed. A new technique is presented and shown to be effective in extracting the qualities of the skin patterning. Methods for analysing this representation of the patterning to quantify the disruption attributable to the lesion are proposed and developed. The combination of these extraction, analysis and disruption evaluation techniques is shown to be effective in relation to both visual assessment of disruption and diagnostic performance.