The future of wavelets in medical image processing

with the accelerated growth of harvesting the capabilities of machine learning and machine vision in the medical field, numerous models and networks have been proposed for handling such a sensitive and complex problems. At the same time, the common problem among the field of medical image processing, whether it be classification, localization or segmentation, is the lack of available datasets. Often, the depth and complexity of the popular and most used networks overpower the small datasets and give insufficient results. To alleviate the problem, we designed a variation of the UNET architecture for a limited dataset based on wavelet transform. The paper focuses on the use of wavelet transform on the input images before they are used in the proposed network. The technique is demonstrated using lumbar spine MRIs. Final outcomes show similar, in some cases even better results using less floating point operations and trainable parameters.