Multimodal image retrieval to support medical case-based scientific literature search

The large amount of available digital biomedical information has a great potential to impact both clinical  practice  and  biomedical  research.  Exploiting  this  potential  requires  efficient  and  effective  mechanisms  to  access  this  information.  This  project  proposal  addresses  the  problem  of  retrieving  relevant  information  from  medical  scientific  literature  sources  to  support  case-based  decision  and  analysis  processes.  This  aim  requires  handling  and  exploiting  the  multimodal nature of scientific biomedical literature through data fusion and machine learning techniques. In addition, the  large  size  of  literature  collections  available  today  enables  a  big  data  approach  to  data  driven  discovery,  requiring  not  only  the  development  of  state  of  the  art  algorithms  for  multimodal  data  analysis  and  machine  learning  on  large  datasets,  but also devising the appropriate tools to guarantee the capacity and scalability of the involved technologies. In this context, we  conceive  the  notion  of  “effectiveness”  in  the  sense  of  empowering  researchers  with  the  capability  of  performing  their  experimental  life  cycles  in  an  agile  and  focused  manner,  either  when  building  their  information  retrieval  systems  or  when  such systems are used for a specific purpose. 

Our  long  term  and  foremost  goal  is  to  build  a  system  for  medical  case-based  information  retrieval  on  large  collections  of  scientific  biomedical  literature.  This  goal  encompasses  two  major  research  problems:  (1)  devising  an  effective  multimodal  representation  strategy  that  captures  the  rich  visual  and  textual  content  of  medical  cases  and  scientific  papers;  (2)  designing and implementing efficient algorithms and processing strategies that can cope with the ever growing collections  of  scientific  papers  and  biomedical  data.  This  includes  addressing  the  scalability  of  both  the  devised  strategies  and  algorithms and the underlying technological substrate supporting them. 

Research and Application Areas

The proposal covers the following research topics included in the call for proposals: 

  • Data  driven  discovery.  The  problem  to  be  addressed  by  the  proposed  project  involves  the  analysis  of  large  collections of scientific papers to automatically find connections between the visual and textual content.  
  • Data fusion. An effective indexing of the content of scientific biomedical papers requires to process both the visual  and textual content in them to build a common multimodal representation. This is in fact a data fusion problem.  
  • Machine learning. Machine learning is at the core of the proposed strategy to automatically build the multimodal  representation.   

The  main  application  area  of  the  proposed  system  is  biomedical  research  and  healthcare. Efficient  and  effective  access  to  the humongous volume of biomedical data that is generated nowadays is a current and challenging research problem with  an important potential impact on these areas. Techniques for content-based retrieval of textual and visual information are  an important component of any solution to this problem. 

Application Impact

Medical images are used to visualize biological structures that are not observable otherwise, such as microscopy images for tissue composition, and x-rays for internal organs. These images serve as evidence of the patient’s health status and allow physicians to make decisions about potential diagnosis and treatments. Also, the patient’s health record brings a set of semi-structured textual information describing basic attributes such as gender and age, as well as expert opinions, recommendations, and even test results. 
When a physician is treating a patient, both data modalities, medical images and textual records, may be used to automatically identify related information from scientific sources. An effective search system can help to process the multimodal query and retrieve relevant documents from a collection of up to date scholarly articles. These results may support the decision making process in medicine by allowing to identify images in papers that illustrate medical cases as well as reports of the latest procedures for treating patients. 
The multimodal search system can be a valuable tool for Evidence-Based Medicine (EBM), which aims to use the best possible evidence for making decisions about the care of individual patients. In that sense, the ability of using the actual health record for finding useful information in scientific archives will help to support the underlying medical reasoning. Then, the process will result in a better understanding of the situation, and therefore, more informed decisions that will impact the quality of life of patients.