Skin Lesion Analysis towards Melanoma Detection Using Deep Learning Network

Skin lesions are a severe disease globally. Early detection of melanoma in dermoscopy images significantly increases the survival rate. However, the accurate recognition of melanoma is extremely challenging due to the following reasons: low contrast between lesions and skin, visual similarity between melanoma and non-melanoma lesions, etc. Hence, reliable automatic detection of skin tumors is very useful to increase the accuracy and efficiency of pathologists. In this paper, we proposed two deep learning methods to address three main tasks emerging in the area of skin lesion image processing, i.e., lesion segmentation (task 1), lesion dermoscopic feature extraction (task 2) and lesion classification (task 3). A deep learning framework consisting of two fully convolutional residual networks (FCRN) is proposed to simultaneously produce the segmentation result and the coarse classification result. A lesion index calculation unit (LICU) is developed to refine the coarse classification results by calculating the distance heat-map. A straight-forward CNN is proposed for the dermoscopic feature extraction task. The proposed deep learning frameworks were evaluated on the ISIC 2017 dataset. Experimental results show the promising accuracies of our frameworks, i.e., 0.753 for task 1, 0.848 for task 2 and 0.912 for task 3 were achieved.


Introduction
Melanoma is the most deadly form of skin cancer and accounts for about 75% of deaths associated with skin cancer [1].Dermoscopy technique has been developed to improve the diagnostic performance of melanoma.Dermoscopy is a noninvasive skin imaging technique of acquiring a magnified and illuminated image of a region of skin for increased clarity of the spots on the skin [2].It enhances the visual effect of skin lesion by removing surface reflection of skin.Deep learning approaches using dermoscopy images to automatically detect melanoma have been proposed in recent research [3].International Skin Imaging Collaboration (ISIC) continuously organized melanoma detection challenges from 2016, which highly promotes the improvements of automatic melanoma detection methods.
In this paper, we proposed two deep learning networks to address the three tasks announced in ISIC 2017 challenges.A multi-scale fully-convolutional residual network is proposed to simultaneously address task 1, i.e. lesion segmentation, and task 3, i.e. lesion classification, and a CNN-based framework is proposed for task 2, i.e. dermoscopic feature extraction.

Methods
In this section, we introduce the deep learning methods for different tasks.

Lesion segmentation and classification (task 1 & 3)
The fully convolutional residual net (FCRN) proposed in our previous work [4] has been extended to simultaneously address the tasks of lesion segmentation and classification.Using the proposed FCRN, we construct a multi-scale deep learning network for skin lesion images.
Data Augmentation.The original ISIC skin lesion dataset contains 2000 images.To enlarge the lesion area for feature detection, we proportionally cropped the center area of each image before augmentation.As the image volumes of different categories widely vary, we accordingly rotated the images belonging to different classes to establish a class-balanced dataset.The dataset augmented with this step is denoted as DR.The images in DR are randomly flipped along x or y-axis to establish another pair dataset, called DM.The two datasets are used to train FCRNs, respectively.
Network Architecture.The flowchart of proposed multi-scale deep learning network is presented in Fig. 1.Two FCRNs trained with datasets using different data augmentation methods are involved.The architecture of FCRN is the same as that introduced in [4].As fully-convolutional network accepts inputs with different sizes, we resize the skin lesion images to two scales, i.e. 300x300 and 500x500.The lesion index calculation unit (LICU) is designed to measure the probabilities for Melanoma, Nevus and Seborrheic keratosis.The reason for using separate FCRN trained on different datasets, i.e.DR and DM, is that we found 'mirror' operation seems to fool the FCRN during training.In our experiments, single FCRN trained on the combination of DR and DM was difficult to converge.The segmentation and classification accuracies on validation set verified our findings, i.e. the separate network provides better segmentation and classification performance than that of single FCRN trained using DR+DM.

Dermoscopic feature extraction (task 2)
We developed a CNN-based approach to address the task 2, i.e. dermoscopic feature extraction.To our best knowledge, this is the first work using deep learning to extract dermoscopic feature from skin lesion images.
Training Data.The ISIC dermoscopic images have four kinds of features, i.e.Network, Negative Network, Streaks and Milia-like Cysts.dermoscopic images were separated by superpixel masks, which contain the feature information of different dermoscopic images.We used the provided algorithm to extract the content of each superpixel and resize them to a uniform size, i.e. 56x56, for our CNN framework.
Network Architecture.Our CNN framework was trained with the patches extracted from superpixel masks.The flowchart of our CNN framework is presented in Fig. 2.
While the blue rectangles represent the convolutional layers, the numbers represent kernel size and number of kernels.Both max pooling and average pooling are used and the network was trained with softmax loss.

Results
We participated all the three tasks announced in ISIC 2017 challenges and achieved promising results, as shown in Table 1, on validation sets.

Conclusion
In this paper, we briefly introduce the deep learning methods proposed for ISIC 2017 challenge.Our approaches achieve promising results on the validation set and have the potential to be the computer-aid diagnosis system for melanoma detection.

Table 1 .
Accuracies on the ISIC 2017 validation set