Dermoscopy Image Classification Based On Deep Active Learning

Malignant melanoma is the deadliest type of skin cancer,and early detection is crucial to improve survival rates.Dermoscopic is an image technique that visualizes the deep level of the skin,and it is widely used to diagnose melanoma.However,even for professional dermatologists,manually checking dermoscopy images is an error-prone and time-consuming task.Therefore,it is very necessary to develop a computer support system for automatic classification to analyze dermoscopy images.Deep learning technology,as an automatic classification system,can reach or exceed the level of some experts in the medical field,but this usually occurs only when the target task has a large number of annotated samples.However,annotation of medical image data is not only tedious,but also consumes a lot of time,which makes annotation cost very high.Therefore,in order to better apply the deep learning technology(maximizing the performance of the model while reducing the cost of medical image annotation)in the classification of melanoma images based on skin cancer,this dissertation systematically analyzes and studies the related problems faced.The main research work and achievements of the thesis are as follows:1.Aiming at the problem of data imbalance,this thesis uses a data enhancement method,and proposes a random grayscale masking data enhancement method on the basis of improving others’ methods,thereby improving the model performance;2.Aiming at the problem of redundant samples in training samples,this thesis proposes and implements the BQU algorithm based on active learning technology.The BQU algorithm essentially integrates deep learning and active learning into one framework(deep active learning framework).Therefore,on the one hand,the three main parts of the deep active learning framework(Build the initial training sample set,query strategy,and update model)are comprehensive research;on the other hand,carry out detailed research on three aspects of Build the initial training sample set,query strategy and update model,and analyze and implement the latest effective technology in the corresponding parts.It is verified by related comparative experiments on the ISIC2016 dermoscopy image data set:(1)The improved enhancement method can be more effectively applied to the classification task of dermoscopy images,and the AUC value can reach 82%,which is nearly 6.3% higher than the effect without data enhancement;(2)Combined with the above research results,when using the BQU algorithm,the improved data enhancement method was used during training.Finally,the BQU algorithm only used 800 of the 900 training images to achieve results comparable to the baseline.It can be seen that the cost of annotation using the BQU method can be reduced by at least 11%.This shows that the method proposed in this thesis can effectively reduce the cost of medical image annotation,and lay a good foundation for the more efficient application of deep learning technology in the medical field.