Study And Application Of Deep Learning Based Classification In Fetal Ultrasound Images And Dermoscopy Images

In the diagnosis of clinical image,the accurate image analysis of a disease is of great important for its treatment.While image interpretation by human experts is limited due to its labor-intensive time-consuming,and subjectivity.To circumvent these problem,automatic image analysis system based on computer algorithms is developed.Recently,deep learning methods,specifically deep convolutional neural network,have established an overwhelming presence in computer aided diagnosis.Compared with traditional machine learning algorithms,deep learning approaches capable of learning from raw data,thus complex image preprocessing and domain-specific feature engineering avoided.State-of-the-art results on numerous visual tasks demonstrated its superiority.However,several challenges still existing in the application of deep learning in medical image analysis.For instance,optimization problem incurred by insufficient training data.Fine-grained variability among different categories images,which pose great difficulty for high level convolutional features.Also,there is huge variations due to images obtained by different acquisition system and conditions.In this study,we conduct research to address the challenges existing in the application of deep network in the classification of fetal ultrasound images and dermoscopy images.Two major kind of algorithms are investigated,i.e.end-to-end trained network,and hybrid model by combing deep representations and local descriptor encoding.Our contributions mainly includes:1.Based on fetal facial ultrasound standard plane images,we proposed a 19 layers deep convolutional network.Data preprocessing,model architecture designing,and training strategy are explored to address the optimization problem of the model under limited training data.We optimized the architecture of our model to enhance the recognition performance by combing a series techniques,i.e.data augmentation,global average pooling layer,and transfer learning based on fine tuning.The extensive experiments on our collected FFSP dataset demonstrated the superiority of our method over the traditional classification models for FFSP recognition.The impressive performance indicated the proposed method are invariant toward the ultrasound image noise,and capable of learning high level representations more efficiently.2.We present a novel framework for dermoscopy image assessment.The proposed framework take advantage of both deep learning method and local descriptors encoding strategy.Specifically,the deep representations of a rescaled dermoscopy image are first extracted via a very deep residual neural network(Res Net)pre-trained on a large natural image dataset.Then these local deep descriptors are aggregated by an orderless visual statistics based on fisher vector(FV)encoding to build a global image representation.Finally,the FV encoded representations are classified for melanoma recognition using a support vector machine(SVM)in order to address the fine-grained variations among different dermoscopy images under limited training data.Also,multi-layers features are fused to further improving the result.We demonstrate the effectiveness of the proposed method on publicly available ISBI 2016 Skin lesion challenge dataset.Comparison with state-of-the-art methods shows the superiority of our method.3.We propose a framework for automatic skin lesion recognition by aggregating selected deep representations extracted from multiple convolutional networks.Compare to the previous deep feature encoding method,several techniques are developed to improve the recognition result.On one hand,color-based data augmentation which modeling the potential distribution of the data,is explored to enlarge the skin data set.On the other hand,we introduce a selective algorithms to extract local deep convolutional features that most likely related to interested lesion object.Also,features from multiple networks are extracted and fused to further improve the performance.All the solutions are combined aiming to address the problem existing in skin lesion image recognition.Such as huge difference in imaging conditions,disturbance incurred by non-related objects.Extensive experiments are performed to investigate the effectiveness of our proposed method on publicly available ISBI 2016 and 2017 Skin lesion challenge dataset,the significant performance improvement indicated its superiority.Aiming to address the potential challenges existing in the classification of fetal ultrasound images and dermoscopy images using deep learning method,we conduct research and propose a series of solution including: 1)design and train an end-to-end network architecture,and give the predication;2)treat deep network as feature extractor,and combine it with traditional machine learning model.Meanwhile,extensive and systematic experiments are conducted to verify the effectiveness of proposed method.The algorithms developed in this work can be a good example for designing other type of automatic clinical images analysis.