Melanoma is a type of skin cancer with the most rapidly increasing incidence. Early detection of melanoma using dermoscopy images significantly increases patients' survival rate. However, accurately classifying skin lesions by eye, especially in the early stage of melanoma, is extremely challenging for the dermatologists. Hence, the discovery of reliable biomarkers will be meaningful for melanoma diagnosis. In recent years, the value of deep learning empowered computer-assisted diagnose has been shown in biomedical imaging-based decision making. However, much deep learning research focuses on improving disease detection accuracy but not understanding the features deep learning use to determine the evidence of pathology. We aim to make sure the features used by deep learning methods are the reasonable clinical features for skin lesions diagnosis, rather than artifacts. Further, we aim to discover new biomarkers, which may not have been included in clinical criteria but do make sense to the dermatologists. Our proposed pipeline can find biomarkers for identifying different lesions. The patterns are agreed with dermatologists. Surprisingly, we find surround skins also can be used as evidence for skin lesion diagnosis, which has not been included in traditional diagnosis rules. The biomarkers discovered from deep learning classifier can be significant and useful to guide clinical diagnosis.