Multimodal image analysis in tissue diagnostics for skin melanoma

Early diagnosis is a corner stone for a successful treatment of most diseases including melanoma, which cannot be achieved by traditional histopathological inspection. In this respect, multimodal imaging, the combination of TPEF and SHG, features a high diagnostic potential as an alternative approach. Multimodal imaging generates molecular contrast, but to use this technique in clinical practice, the optical signals must be translated into diagnostic relevant information. This translation requires automatic image analysis techniques. Within this contribution, we established an analysis pipeline for multimodal images to achieve melanoma diagnostics of skin tissue. The first step of the image analysis was the pre-treatment, where the mosaicking artifacts were corrected and a standardization was performed. Afterwards, the local histogram-based first-order texture features and the local gray-level co-occurrence matrix (GLCM) texture features were extracted in multiple scales. Thereafter, we constructed a local hierarchical statistical model to distinguish melanoma, normal epithelium, and other tissue types. The results demonstrated the capability of multimodal imaging combined with image analysis to differentiate different tissue types. Furthermore, we compared the histogram and the GLCM-based texture feature sets according to the Fisher's discriminant ratio (FDR) and the prediction of the classification, which demonstrated that the histogram-based texture features are superior to the GLCM features for the given task. Finally, we performed a global classification to achieve a patient diagnostics with the clinical diagnosis as ground truth. The agreement of the prediction and the clinical results demonstrated the great potential of multimodal imaging for melanoma diagnostics. Multimodal imaging was applied for tissue diagnostics of skin melanoma with the help of automatic image analysis procedures. Normal, melanoma, and other skin tissues were classified based on multiscale pixel-wise texture features calculated from first-order histogram and Gray-Level Co-Occurrence Matrix of the multimodal images. Texture features based on first-order histogram were superior to those based on Gray-Level Co-Occurrence Matrix with respect to computational complexity, classification accuracy, and Fisher's discriminant ratio.