Research On Medical Image Denoising And Lung Cancer Detection Using Convolutional Neural Networks

Medical images are tremendously important element in medical practice in this day.The recent development of various imaging system that produce images with riches of information has greatly enhanced the applicability and significance of medical images.Although the imaging system that produces medical images are showing improvement through time,no image generated are free of noise.Noise covers and reduces the visibility of certain features within the image,which compromise an image quality and affects a subsequent medical image application.So,denoising can help to enhance the quality of an image attacked by noise.On the other hand,due to its immense importance in a subsequent medical image investigation,medical image detection method has obtained a great consideration in medical image applications.Medical image nature is inconsistent across a patient.For example,physically,the size,shape,and the location of a tissue is variable from patient to patient,even for an individual patient,within an image.Similarly,their correlations are varying,and they are relying on both local and global features.Due to this variation,detection of a medical image is challenging and is often missed by radiologists.Therefore,in addition to the denoising task,medical image detection method is another area of interest to investigate.This work is focused on medical image denoising and detection task.Denoising is the first and vital step in medical image diagnosis,whereas detection is essentially important for subsequent processes such as for segmentation,registration,and diagnosis in general.The first part of this thesis introduces two medical image denoising techniques using deep convolutional neural network(CNN).Network depth is very crucial for a low level task.However,with the increase of the network depth,there is a problem of gradient vanishing which hinder the performance of the network.To solve this,the first denoising technique is based on denoising convolutional neural network with residual learning approach(DCBN-Netr).We design a feed forward denoising CNN by taking into consideration the network depth,learning algorithm and normalization method.Accordingly,residual learning is adopted as learning approach and batch normalization is used as regularization method for medical image denoising.More specifically,unlike most of the other CNN based image denoising approaches which directly learn the latent clean images,the residual learning is adopted to learn the noise by removing latent clean image,where the noise free images are obtained by subtracting the estimated noise from the noisy image.Employing residual learning and batch normalization with deep network speed up the training process as well as boost the denoising performance.Experimental results demonstrate that the proposed model improves the average reconstructed image quality by 0.1d B to 0.9d B.The second technique proposes a novel feature map smoothing constrained denoising network(FMSCD Net)by introducing feature map smoothing constraint in the intermediate CNN layers.First,we analyze the statistical properties of the feature maps of intermediate CNN layers for clear medical images and verify that the intermediate feature map has smoothing property(e.g.,their gradient features are sharp and zero values are in majority).However,these smoothness properties are violated from feature maps of noisy medical images because of the noise,and during denoising via CNN's,we have observed that the quality of denoised images is somehow deteriorated.In order to keep the smoothness properties of the feature maps of CNN's for the degraded images,a novel smoothing constraint is introduced.Given a feature map,each of its values smoothness property is retained by weighting their local neighborhoods.The weights are computed based on the similarity of two consecutive receptive field.Second,by applying this smoothing constraint,an end to end FMSCD Net which directly approximate the latent clean image is designed.The constrained feature map helps to train deeper architecture and improves the reconstructed image quality.Finally,we integrate residual learning along with batch normalization into the FMSCD-Net to improve the model performance.Experimental results reveal that our approach has improved the average reconstructed image quality by 0.01 to 1.1d B.The second part of this thesis introduces two medical image detection approaches using multi path CNN's.Taking a CT scan image of lungs,both models are applied to lung cancer detection problem.Lung cancer is the leading cause of death among cancer related death.Like other cancers,the finest solution for lung cancer diagnosis and treatment is early screening.Since the presence of pulmonary nodules in a CT scan image of a lung does not absolutely specify cancer,the morphology of nodules such as shape,size,and contextual information has a sophisticated relationship with cancer,the screening of lung cancer needs a careful investigation on each suspicious nodule and integration of information of all nodules.To solve this problem in the first model,we propose a multi-path convolutional neural network(MP-CNN)used to detect lung cancer.First,in the pre-processing step,the suspicious nodules are generated with the modified version of U-Net and then the generated nodules become input data for the model.The proposed model is a multi-path CNN which exploits both local features as well as more global features simultaneously to automatically detect lung cancer.To this end,the model used three paths,each path employed different receptive field size which helps to model distant dependencies.Then,to further upgrade our model performance,we concatenate features from the three paths.Finally,we also introduce a retraining phase system that permits us to tackle difficulties related to the imbalance of image labels.The model achieved87.8% detection accuracy.The second lung cancer model is a “denoising first detection” two-path convolutional neural network(DFD-Net)which perform both denoising and detection designed with efficient feature fusion strategy.First,a residual learning denoising network(DCBN-Netr)is employed to remove noise from the noisy version of the CT scan image of the lung.It is composed of convolution and batch normalization layers which construct residue instead of the latent original image.Second,the two-path convolutional neural network is employed to handle the detection task.It uses the denoised image by DCBN-Netras an input to detect lung cancer.The two paths intended to model local features and global features.For this purpose,each path used different receptive field size which aids to model distant dependencies.Third,to further polish the model performance,different from the conventional feature concatenation approaches which directly concatenate two sets of features from different CNN layers,discriminant correlation analysis is introduced to concatenate more representative features.We found that this type of model easily first reduce noise in an image,balances the receptive field size effect,affords more representative features,and makes our model more adaptable to the variability of shape and size among nodules.Experimental result shows that the model achieved 88.3%detection accuracy.