With the development of medical image equipment, the Computed tomography (CT) scan has greatly increased the detectable area of the human body and improved the detection sensitivity and accuracy of disease monitoring. However, the unbalanced distribution of HU intensity values lowers accuracy and sensitivity of automatic analysis and segmentation based on CT images. As such, we proposed a redistributing method to make the distribution even while ensuring that similar intensities of CT images reflect the same tissue after processed. Our experiments based on the 3D brain tumor CT images proved that this method can significantly speed up the convergence and improve segmentation accuracy as compared to conventional preprocessed methods.

With regard to pulmonary nodule detection, due to the similar texture and shape as particular tissues, it is difficult for Computer-Aided Detection (CAD) system in detecting pulmonary nodule with both high accuracy and sensitivity. To address this problem, we design a 3D automated pulmonary nodule detection where a auxiliary 3D generative adversarial network is embedded. This well-trained auxiliary component that fully learns volumetrically contextual information of nodule and non-nodule structure, is exploited for each input sample of detection model to generate a derivative which only preserve background context by removing all the nodules. By learning the feature contrast between each input and its derivative, our detection model achieves competitive performance to state-of-the-art approaches for the pulmonary nodule detection task.

MR technique is prevalent for doctor to diagnose and assess glioblastomas which are the most lethal form of brain tumors. Although Convolutional Neural Networks (CNN) has been applied in automatic brain tumor segmentation and is proved useful and efficient, traditional one-pathway CNN architecture with convolutional layers and max pooling layers has limited receptive fields representing the local context information. Such mindset in traditional CNN may dismiss useful global context information. In this paper, we design a two-pathway model with average and max pooling layers in different paths. Besides, 1 x 1 kernels are followed input layers to add the non-linearity dimensions of input data. Finally, we combine the CNN architecture with fully connected CRF(FCRF) as a mixture model to introduce the global context information to optimize prediction results. Our experiments proved that the mixture model improved segmentation and labeling accuracy. (C) 2018 Elsevier Inc, All rights reserved.

Brain tumor segmentation on MR images has significant clinical meaning due to glioblastomas which are the most lethal form of these tumors. Compared to manual segmentation, automatic segmentation system is superior in timesaving and experience-insensitivity for doctors during clinical practice. However, its inherent contradiction is not addressed yet. i.e. imbalance of multi-class of different brain tissues. As such, we proposed a multi-class focal loss to make the loss function emphasis on bad-classified voxels in MR images. Our experiments based on the 3D UNet model proved that this method can significantly improve labeling and segmentation accuracy as compared to other loss layers.