Classification of skin diseases based on weakly supervised fine-grained methods
Article title
Skin Lesion Classification Using Weakly-supervised Fine-grained Method
Article Source
2020 25th International Conference on Pattern Recognition (ICPR)
Author motivation
There are many existing works on skin disease classification based on convolutional neural networks (CNN), but few involve the similarity between different lesions. For example, melanoma (malignant) and moles (benign) are very similar in appearance. , but it is difficult for neural networks to distinguish.
Author's ideas
Spatial attention and channel attention are used to enrich feature maps and guide the network to focus on salient regions in a weakly supervised manner. Preprocessing methods such as component elimination and data augmentation are also used to improve accuracy.
Network Architecture
For an input image, components are first eliminated to avoid affecting subsequent feature extraction. Secondly, the network determines whether the image needs to be enhanced based on the image label. The authors performed data augmentation only on melanoma and seborrheic keratosis because there are much more data on moles than the other two. The network is then split into two branches. The upper branch uses CNN to extract global features from the complete image. The lower branch is the DRPM module, which outputs the area with the largest amount of information, adjusts the image size and sends it to CNN to extract local features. The images obtained by the last two branches are spliced together to obtain final result
The DRPM module includes four modules:
- Attention module:
"Bold part" explanation
Eliminate parts : first convert the three-channel image into a one-channel grayscale image, then perform morphological transformation on the image to eliminate noise, then set a 17*17 kernel to perform blackhat operation to highlight the outline of the part, and next perform a threshold operation, using to extract objects in images to distinguish background and parts. The last step is the inpainting operation, which restores the damage to the image caused by the above steps. The inpainting operation is a binary image obtained based on the threshold operation. The pixels near the area boundary are used to reconstruct the pixels of the removed part. The following figure is an example before (a) and after (b) the component is eliminated:
Data augmentation : The goal of data augmentation is to increase the number of categories with fewer samples so that each category is relatively balanced. There are many methods of data enhancement, such as: color adjustment, random cropping, random flipping, and radial transformation. The following is the rendering of data enhancement:
