|dc.description.abstract||Clinical analysis and medical diagnosis of diverse diseases adopt medical imaging techniques to empower specialists to perform their tasks by visualizing internal body organs and tissues for classifying and treating diseases at an early stage. Content-Based Image Retrieval (CBIR) systems are a set of computer vision techniques to retrieve similar images from a large database based on proper image representations. Particularly in radiology and histopathology, CBIR is a promising approach to effectively screen, understand, and retrieve images with similar level of semantic descriptions from a database of previously diagnosed cases to provide physicians with reliable assistance for diagnosis, treatment planning and research.
Over the past decade, the development of CBIR systems in medical imaging has expedited due to the increase in digitized modalities, an increase in computational efficiency (e.g., availability of GPUs), and progress in algorithm development in computer vision and artificial intelligence. Hence, medical specialists may use CBIR prototypes to query similar cases from a large image database based solely on the image content (and no text). Understanding the semantics of an image requires an expressive descriptor that has the ability to capture and to represent unique and invariant features of an image. Radon transform, one of the oldest techniques widely used in medical imaging, can capture the shape of organs in form of a one-dimensional histogram by projecting parallel rays through a two-dimensional object of concern at a specific angle. In this work, the Radon transform is re-designed to (i) extract features and (ii) generate a descriptor for content-based retrieval of medical images. Radon transform is applied to feed a deep neural network instead of raw images in order to improve the generalization of the network. Specifically, the framework is composed of providing Radon projections of an image to a deep autoencoder, from which the deepest layer is isolated and fed into a multi-layer perceptron for classification. This approach enables the network to (a) train much faster as the Radon projections are computationally inexpensive compared to raw input images, and (b) perform more accurately as Radon projections can make more pronounced and salient features to the network compared to raw images. This framework is validated on a publicly available radiography data set called "Image Retrieval in Medical Applications" (IRMA), consisting of 12,677 train and 1,733 test images, for which an classification accuracy of approximately 82% is achieved, outperforming all autoencoder strategies reported on the Image Retrieval in Medical Applications (IRMA) dataset. The classification accuracy is calculated by dividing the total IRMA error, a calculation outlined by the authors of the data set, with the total number of test images. Finally, a compact handcrafted image descriptor based on Radon transform was designed in this work that is called "Forming Local Intersections of Projections" (FLIP).
The FLIP descriptor has been designed, through numerous experiments, for representing histopathology images. The FLIP descriptor is based on Radon transform wherein parallel projections are applied in a local 3x3 neighborhoods with 2 pixel overlap of gray-level images (staining of histopathology images is ignored). Using four equidistant projection directions in each window, the characteristics of the neighborhood is quantified by taking an element-wise minimum between each adjacent projection in each window. Thereafter, the FLIP histogram (descriptor) for each image is constructed. A multi-resolution FLIP (mFLIP) scheme is also proposed which is observed to outperform many state-of-the-art methods, among others deep features, when applied on the histopathology data set KIMIA Path24. Experiments show a total classification accuracy of approximately 72% using SVM classification, which surpasses the current benchmark of approximately 66% on the KIMIA Path24 data set.||en