Paper Overview
This paper proposes a method based on the depth of learning to achieve automatic detection of melanoma lesions and segmentation. The method comprises an enhanced encoding - decoding network for extracting a feature of the encoding sub-network and network connection decoder sub-network path through a series of hops (skip pathway), characterized in that the mapping of both the (feature maps) closer . Further, the system uses a multi-stage method, the first pixel using the softmax-class classifier classifier, classifier after use injury (Lesion Classification) classification analysis based on the results of skin lesions where pixel level classification.
Brief introduction
This paper uses a single convolution depth neural network (DCNN) in all processes. The method uses an enhanced depth supervision encoding - decoding (encoder-decoder) network image feature extraction. The network may be extracted from a multi-stage process by which complex characteristic lesion image, wherein the coding stage of learning the general appearance, including hair may affect the positioning information and the lesion area, the lesion boundary feature learning decode stage. The difference between network and existing methods of paper proposes that the following three aspects:
(1) through a series of hops path encoding and decoding sub-network the sub-network are connected together, thereby enhancing the characteristics of learning ability and ability to extract characteristics of the network;
(2) multi-scale design on each network path hops system, skin lesions processing images of different sizes;
(3) a method using Lesion-classifier to the pixel level hierarchy into the skin lesions of melanoma and non-melanoma.
The algorithm can be used with limited training image data set and detect melanoma in the case of limited computing resources to meet the needs of real-time clinical practice.
Self-organization network
Figure 1 shows the pre-processing to extract features from the image, then the pixel cell classification (Pixel-wise Classification), and finally to the main stages of lesion classification. Skin damage includes image data set, the encoding - decoding network, damage the Softmax classifier and a classifier. The image data input to the encoding set - first preprocessed before decoding the network. Decoding network is trained - convolutional coding using the skin depth mirror image (dermoscopic images) tagged. Input into the first encoder sub-networks, sub-networks and then fed to the decoder for feature extraction. Further a module loss function dice (dice loss function) and the combined softmax classifier for the image pixel level classification and recognition of melanoma sensitive areas.
1. Data pre-processing and image enhancement
In the preprocessing stage using a Gaussian filter to separate image noise. After adjusting the image to achieve the same ratio and resolution. By calculating the mean pixel value data and the normalized standard deviation of the image normalization process. Further, a simple method in the random rotation of the enhancement process to improve performance. By linearly mapping intensity to a [- 0.5, 0.5] to the pixel intensity range centered around 0, so as to provide numerical stability during the training.
2. ROI feature extraction and recognition
As used herein, a coding depth learning framework with - decoder network used to extract features, wherein the encoding section and decoding by the decoding section 5 blocks respectively (Figure 2).
The encoder main learning and general appearance of the input image information to capture the positioning information, each block of 3 × 3 convolution of two layers, a cell layer maximum (max-pooling) and one for non-feature extraction ReELU composed of a linear activation function. Wherein the maximum cell layer eliminates redundant features, enable the computing time is minimized. Convolution and ReLU layer activation function in the learning mode lesion image and the end pixel system training process. The encoder and decoder sub-networks are connected by a series of hops path (short jump convolutional network and the network).
Decoder portion is also composed of five units, each unit consists of two layers and a convolution of the sample layer. In the decoding section, the use of layer 2 × 2 convolution output of a previous sample block is the nearest neighbor. And then outputs it to the appropriate level of the encoder connected portion.
3. pixel-level classification of skin damage & Classification
Encoding - decoding network after training using the pixel disease and boolean tag and obtain a set of high-dimensional feature, is implemented using two classifiers: pixel-level classification and classification of skin damage. First, the high dimensional feature softmax classifier is fed in order to predict the specific class for each pixel is whether the melanoma cells. Thereafter, the skin lesions due to the low contrast image and the presence of the surrounding skin, among other factors, the dice loss function using a feedback network to improve network performance prediction. Finally obtained by the detecting portion classifier damage results.
Conclusion
This paper proposes a method of detecting melanoma convolutional network based on depth divided. The system was evaluated in two open skin lesion image data set. The overall accuracy of the system and dice ISIC 2017 challenge dataset coefficient of 95% and 92%, respectively, of PH2 images and data sets of dice coefficient accuracy of 95% and 93%, respectively.
