Medical Image Retrieval for Augmenting Diagnostic Radiology

Ashery Mbilinyi
PhD Thesis
Appears in
PhD Thesis, Department of Mathematics and Computer Science
University of Basel, Switzerland

Even though the use of medical imaging to diagnose patients is ubiquitous in clinical settings, their interpretations are still challenging for Radiologists. Many factors make this interpretation task difficult, one of which is that medical images sometimes present subtle clues yet are crucial for diagnosis. Even worse, on the other hand, similar clues could indicate multiple diseases, making it challenging to figure out the definitive diagnoses. To help radiologists quickly and accurately interpret medical images, there is a need for a tool that can augment their diagnostic procedures and increase efficiency in their daily workflow. A general-purpose medical image retrieval system can be such a tool as it allows them to search and retrieve similar cases that are already diagnosed to make comparative analyses that would complement their diagnostic decisions. This thesis aims to contribute to the development of such a system by proposing approaches to be integrated as modules of a single system, enabling it to handle various information needs of radiologists and thus augment their diagnostic processes during the interpretation of medical images.

We have mainly studied the following retrieval approaches to handle radiologists’ different information needs; i) Retrieval Based on Contents, ii) Retrieval Based on Contents, Patients’ Demographics, and Disease Predictions, and iii) Retrieval Based on Contents and Radiologists’ Text Descriptions. For the first study, we aimed to find an effective feature representation method to distinguish medical images considering  their semantics and modalities. To do that, we have experimented with different representation techniques based on handcrafted methods (mainly texture features) and deep learning (deep features). Based on the experimental results, we propose an effective feature representation approach and deep learning architectures for learning and extracting medical image contents. For the second study, we present a multi-faceted method that’s complements image contents with patients demographics and deep learning-based disease predictions making it able to identify similar cases accurately considering the clinical context the radiologists seeks.

For the last study, we propose a guided search method that combines an image with a radiologist’s text description to guide the retrieval process. This method guarantees that the retrieved images are suitable for the comparative analysis to confirm or rule out initial diagnoses (the differential diagnosis procedure). Furthermore, our method is based on a deep metric learning technique and is better than traditional  content-based approaches that rely on image features only and, thus, sometimes retrieve random images that are insignificant.

Staff members