Early diagnosis of melanoma, which is considered to be one of the deadliest skin cancers, via medical imaging can significantly improve the course of the disease. However, expert assessments are subjective and open to errors due to large variations in dermoscopy images. To cope with this problem, a two-stage framework is proposed for the detection of melanoma in dermoscopic images. Firstly, by eliminating the presence of natural structures such as veins or hair and the variations in the pattern region, segmented images are obtained from raw image data with the help of pixel-wise image processing techniques. The second part of this framework is the recognition stage of the skin lesions by Pre-Trained Deep Networks (PTN). By using segmented input images, the PTN classifiers are optimized with hyperparameters and classification is performed. The performance of the framework was evaluated separately before and after segmentation for six PTN used in this study. An improvement of 5.1% in the Accuracy metric and 7.7% in the Jaccard metric have been observed in the lesion recognition process with the segmentation framework according to the averages of network performances.