Skin Cancer Segmentation Using Deep Fully Convolutional Networks

Detecting Skin lesions in dermoscope images is very difficult due to indefinite lesion boundaries. A type of cancer known as malignant melanoma is increasing all over the world. The reported cases each year due to this cancer is alarming. Immunotherapy and Radiotherapy are advance techniques that use high radiations to eliminate cancer cells. It is very important to diagnose this cancer in the early stages because the five-year survival rate is 99% according to the latest statistics [1]. 

Dermatologists use dermoscopic tool to explore the skin lesion having pigments according to the morphological (particularly structural) features. So it is diagnosed very accurately by dermoscopy rather than ABCD criteria, still, computerized analysis techniques are introduced because the exact detection and verification take a lot of time in dermoscopy. As human skin is different in color and texture therefore, detecting melanoma in the skin is very difficult because of its appearance in color, shape, and size in human skin.

In the paper, a model is proposed handling dermoscopic images in different learning conditions using Deep CNN based novel framework to automatically segment skin lesions in dermoscopic images. The Fully Convolutional Network is introduced in this paper utilizes 20 layers of deep CNN without using any pre-trained model, which further extends the convolutional process in the whole image predicting the segmentation mask.

In the recent years, many clustering, region splitting and merging, and thresholding based algorithms were introduced to tackle these challenges. The drawbacks of these algorithms is debated. As these techniques divide the input image into small patches and then feed into CNN to check each patch whether it is on background or foreground, providing only limited contextual information, for each patch only gives a local region of an image. If patch size is increased then we may lose fine details. If we use sliding window that has highly overlapped patches to get the local and global contextual information in segmentation but this is computationally very expensive and inefficient.

To ensure efficient and effective learning with a small set of data for training, a set of techniques are introduced. As the foreground and background pixel difference is unequal in the image segmentation therefore, cross entropy is used as a loss function reducing the cross entropy bias in the background because the lesion occupies very small area in the dermoscopic image. The solution proposed to this problem is to define a novel loss function based on Jaccard distance which is used to stop sample re-weighting.

There are 2 path ways in this model, starting from convolutional path combining the contextual information using convolution and pooling and then following deconvolutional path recovering the full image resolution using the deconvolution and up-sampling as shown in the following figure [2]. The architecture of the proposed model consist of 19 layers, using RELUs as activation function in each Cov/de-Cov layer. Sigmoid is used as an activation function for the output layer. The deconvolutional layer is recovering the image details at different level capturing the fine details and global information for the malignant tumor segmentation.