A dual-tree complex wavelet transform based convolutional neural network for human thyroid medical image segmentation

This research proposes a novel dual-tree complex wavelet transform based Convolutional Neural Network (WCNN) to perform organ tissue segmentation from medical images. Accurate and efficient segmentation on the medical image of human organ is a critical step towards disease diagnosis. For medical image segmentation tasks, conventional Convolutional Neural Networks (CNNs) are: 1) inclined to ignore crucial texture information of the image due to the limitations of typical pooling approaches, and 2) insufficiently robust to noise. To overcome the obstacles, a spectral domain transformation technique is adopted in the CNN. Specifically, a dual-tree complex wavelet pooling layer is concatenated to the traditional pooling process in a CNN. By using wavelet decomposition, the image becomes scalable in the spatial direction, allowing accurate recognition of textures. The WCNN decomposes the image into a number of wavelet subbands, and reduces noisy data by filtering out high-frequency subbands. The performance of WCNN is tested on standard image classification datasets, and applied for human thyroid optical coherence tomography (OCT) image segmentation. Compared to the traditional CNNs using max pooling, experimental results demonstrate that the WCNN approach obtains outstanding consistency and accuracy in the image segmentation domain.