Skin Cancer Recognition using Deep Residual Network

Skin cancer is a type of disease in which group of cells grow abnormally forming malignant melanoma invading the surrounding tissue. Skin cancer is the third most common and the major cause of non-accidental death in human among the ages of 20 -39 [1]. Malignant Melanoma is one of the most dangerous cancers and can be fatal if untreated. They create changes in skin surface and color, however early discovery of these progressions can be cured for around 90% of the cases. In 2018, American Academy of dermatology has evaluated that there will be 91,270 new instances of melanoma in the United States and 9,320 deaths from skin cancer. Around 132,000 new instances of skin cancer are analyzed worldwide every year, as per the WHO (World Health Organization).

Digital dermoscopy is one of the most widely used non-invasive, cost effective imaging tool to identify melanomas in patients. It helps in prior diagnosis and allows doctors to give proper treatment that can improve the survival rate.

For this we have proposed an end to end solution that would require only an image as an input. We have followed Deep ResNet approach inspired by [10]. Our aim is to make it easy for people to communicate using the model. There are numerous people who use the ASL around the world. A vision-based approach to our solution attempts to reduce the requirement of human translators and increase dependency on the user’s phone for translation.

Traditionally, skin cancers were diagnosed by analyzing images for nodule development in the body. Manual diagnosis is often laborious, time consuming and may cause inter observer variability. With large volume of medical database, this process of analyzing the images with the help of doctors is hardly reliable [3]. CAD systems help in analyzing various forms of cancer [2] by aiding the doctors as a second opinion reader.

Skin cancers can be nonmelanoma or melanoma. Non-melanoma skin cancers include: Basal cell carcinomas and Squamous Cell Carcinomas. Nonmelanoma types are rarely lethal, on the other hand melanoma skin cancer are lethal. If melanoma is not treated early and removed in time, it can penetrate and grow deep into the skin, it can also spread to other parts of the body. Effects of a late diagnosis of skin cancer can be very significant in terms of personal health.

Melanomas are easily misdiagnosed at an early stage, because it can be confused with benign entities. The diagnosis of melanoma is difficult because of the variability in clinical appearance and an absence of pigmentation. Melanoma can mimic a scar or different tumors [4,5]. Computer Aided diagnosis system extensively began in early 1980s, significant research has been done in using CAD systems for medical image analysis.

In [6] By implementing four step methods, in the first step preprocessing on the images has been done to remove the noisy artifacts like skin hair and light reflection. Second and third step is patch extraction and CNN, a window is patched around the pixels and after that fed it to CNN. CNN analyze the patch from Local texture and General structure. These fully connected layers’ output has been labelled as 0 or 1 in post processing.

While in [7] Lumix SZ1 camera has been used to capture 166 tumor images. These images after then transformed to generate 5000 images by tuning the contrast. Convolutional Neural Network VGG-16 was applied to train on those images. This train network creates feature vectors which in term is used to train Support Vector Machine (SVM) classifier. SVM predicts the location of tumor with the recognition rate of 89.5%.

Processing Units (GPUs), Convolutional Neural Networks (CNNs) have aided in significant progress in computer vision. [8] won the ImageNet challenge in 2012 using a CNN processing on GPU. Medical image diagnosis has been one of the key drivers in the past for leaps in computer vision, the U-Net architecture by [9] which introduced skip connections had performed extremely well on medical images and small size datasets. Classifying images have been hard especially at cellular level, the one which are taken at 100x zoom. [10] Residual learning framework eases the training of deep neural networks. Residual Network has depth up to 152 layers which are 8 time deeper than VGG nets, with lower complexity and by the error margin of 3.57% on ImageNet test set.

Visualizing the layers in a CNN help in understanding the features that influence the algorithm in reaching on a decision. It helps in exploring the reasons behind the failure of the model. The Grad-Cam approach by [11] uses gradients of any class flowing through the final layer to produce a coarse map of highlighting the activation

…(Full text truncated)…