Employing New Hybrid Adaptive Wavelet-Based Transform and Histogram Packing to Improve JP3D Compression of Volumetric Medical Images

The primary purpose of the reported research was to improve the discrete wavelet transform (DWT)-based JP3D compression of volumetric medical images by applying new methods that were only previously used in the compression of two-dimensional (2D) images. Namely, we applied reversible denoising and lifting steps with step skipping to three-dimensional (3D)-DWT and constructed a hybrid transform that combined 3D-DWT with prediction. We evaluated these methods using a test-set containing images of modalities: Computed Tomography (CT), Magnetic Resonance Imaging (MRI), and Ultrasound (US). They proved effective for 3D data resulting in over two times greater compression ratio improvements than competitive methods. While employing fast entropy estimation of JP3D compression ratio to reduce the cost of image-adaptive parameter selection for the new methods, we found that some MRI images had sparse histograms of intensity levels. We applied the classical histogram packing (HP) and found that, on average, it resulted in greater ratio improvements than the new sophisticated methods and that it could be combined with these new methods to further improve ratios. Finally, we proposed a few practical compression schemes that exploited HP, entropy estimation, and the new methods; on average, they improved the compression ratio by up to about 6.5% at an acceptable cost.

Applying 3D U-Net Architecture to the Task of Multi-Organ Segmentation in Computed Tomography

The achievement of high-precision segmentation in medical image analysis has been an active direction of research over the past decade. Significant success in medical imaging tasks has been feasible due to the employment of deep learning methods, including convolutional neural networks (CNNs). Convolutional architectures have been mostly applied to homogeneous medical datasets with separate organs. Nevertheless, the segmentation of volumetric medical images of several organs remains an open question. In this paper, we investigate fully convolutional neural networks (FCNs) and propose a modified 3D U-Net architecture devoted to the processing of computed tomography (CT) volumetric images in the automatic semantic segmentation tasks. To benchmark the architecture, we utilised the differentiable Sørensen-Dice similarity coefficient (SDSC) as a validation metric and optimised it on the training data by minimising the loss function. Our hand-crafted architecture was trained and tested on the manually compiled dataset of CT scans. The improved 3D UNet architecture achieved the average SDSC score of 84.8 % on testing subset among multiple abdominal organs. We also compared our architecture with recognised state-of-the-art results and demonstrated that 3D U-Net based architectures could achieve competitive performance and efficiency in the multi-organ segmentation task.

Semantic Segmentation of 3D Medical Images with 3D Convolutional Neural Networks

A neural network is a mathematical model that is able to perform a task automatically or semi-automatically after learning the human knowledge that we provided. Moreover, a Convolutional Neural Network (CNN) is a type of neural network that has shown to efficiently learn tasks related to the area of image analysis, such as image segmentation, whose main purpose is to find regions or separable objects within an image. A more specific type of segmentation, called semantic segmentation, guarantees that each region has a semantic meaning by giving it a label or class. Since CNNs can automate the task of image semantic segmentation, they have been very useful for the medical area, applying them to the segmentation of organs or abnormalities (tumors). This work aims to improve the task of binary semantic segmentation of volumetric medical images acquired by Magnetic Resonance Imaging (MRI) using a pre-existing Three-Dimensional Convolutional Neural Network (3D CNN) architecture. We propose a formulation of a loss function for training this 3D CNN, for improving pixel-wise segmentation results. This loss function is formulated based on the idea of adapting a similarity coefficient, used for measuring the spatial overlap between the prediction and ground truth, and then using it to train the network. As contribution, the developed approach achieved good performance in a context where the pixel classes are imbalanced. We show how the choice of the loss function for training can affect the nal quality of the segmentation. We validate our proposal over two medical image semantic segmentation datasets and show comparisons in performance between the proposed loss function and other pre-existing loss functions used for binary semantic segmentation.

Lossless Compression of Medical Image Sequences Using a Resolution Independent Predictor and Block Adaptive Encoding

The proposed block-based lossless coding technique presented in this paper targets at compression of volumetric medical images of 8-bit and 16-bit depth. The novelty of the proposed technique lies in its ability of threshold selection for prediction and optimal block size for encoding. A resolution independent gradient edge detector is used along with the block adaptive arithmetic encoding algorithm with extensive experimental tests to find a universal threshold value and optimal block size independent of image resolution and modality. Performance of the proposed technique is demonstrated and compared with benchmark lossless compression algorithms. BPP values obtained from the proposed algorithm show that it is capable of effective reduction of inter-pixel and coding redundancy. In terms of coding efficiency, the proposed technique for volumetric medical images outperforms CALIC and JPEG-LS by 0.70 % and 4.62 %, respectively.