The Impact of Replacing Complex Hand-Crafted Features with Standard Features for Melanoma Classification Using Both Hand-Crafted and Deep Features

Abstract

Melanoma is the deadliest form of skin cancer and it is the most rapidly spreading cancer in the world. An earlier detection of this kind of cancer is curable, hence earlier detection of melanoma is pre-eminent. Because of this fact, a lot of research is being done in this area especially in automatic detection of melanoma. In this paper, we are proposing an automatic melanoma detection system which utilize a combination of deep and hand-crafted features. We analyzed the impact of using a simpler and standard hand-crafted feature, in place of complex usual hand-crafted features eg. shape, texture, diameter, or some custom features. We used a convolutional neural network (CNN) known as deep residual network (ResNet) to extract the deep features and utilized the scale invariant feature descriptor (SIFT) as the hand-crafted feature. The experiments revealed that, combining SIFT did not improve the accuracy of the system, however we obtained higher accuracy than state of the art methods with our deep only solution.

Analyzing Digital Image by Deep Learning for Melanoma Diagnosis

“…The easiest and most common method to reduce overfitting on image data is to artificially enlarge the dataset using label-preserving transformations (e.g., [25,4,5]). We employ two distinct forms of data augmentation, both of which allow transformed images to be produced from the original images with very little computation, so the transformed images do not need to be stored on disk.…”

“…The first form of data augmentation consists of generating image translations and horizontal reflections. We do this by extracting random 224 × 224 patches (and their horizontal reflections) from the 256×256 images and training our network on these extracted patches 4 . This increases the size of our training set by a factor of 2048, though the resulting training examples are, of course, highly interdependent.…”

“…Specifically, we perform PCA on the set of RGB pixel values throughout the ImageNet training set. To each training image, we add multiples of the found principal components, 4 This is the reason why the input images in Figure 2 to the second idea of the proposed architecture: judiciously applying dimension reducprojections wherever the computational requirements would increase too much otherwise. sed on the success of embeddings: even low dimensional embeddings might contain a lot tion about a relatively large image patch.…”

“…Many previous techniques in dermatological computer aided classification has lacked the generalization capability of medical practitioners due to insufficient data and require extensive preprocessing [2], lesion segmentation [6,7] and extraction of specific features before classification [4,16]. Several learning-based methods were applied form melanoma classification [13,3,15] using the special features of this kind of lesion.…”

Skin Lesion Classification by Ensembles of Deep Convolutional Networks and Regularly Spaced Shifting

“…In addition to this, training datasets lack sufficient quality in the sense of homogeneity in the acquisition procedure and nonexpected objects present in the image, making it necessary to carry out several preprocessing steps [5] and segment the region of interest [6], [7]. Moreover, another commonly used technique is the extraction of features that are used then to improve the classification rate [8], [9]. The use of specific features extracted from the melanoma images was widely used to develop classification models [10]- [12], although the main inconvenience of these approaches is the requirement of specific expertise to extract the adequate features and the high quantity of time necessary to select the most appropriate.…”