Towards an automatic bag-of-features model for the classification of dermoscopy images: The influence of segmentation

The classification of skin lesions in dermoscopy images depends on three critical steps: i) lesion segmentation, ii) feature extraction and iii) feature classification. Lesion segmentation plays an important role since segmentation errors may jeopardize the other two steps, leading to erroneous decisions. This paper studies the robustness of a skin lesion classifier based on a Bag-of-features approach in the presence of segmentation errors. We compare the performance achieved by the system using an automatic segmentation algorithm with the performance obtained using manual segmentation provided by a specialist. We observe a degradation of the system accuracy by 8% when automatic segmentation is used. We also show that these results can be improved if manually segmented images are used in training phase, keeping a fully automatic solution during the testing phase.