How do you train a computer to effectively retrieve medical images? Researchers at Case Western Reserve University in Cleveland, Ohio claim some success as part of an effort to build a large-scale medical image retrieval system for consumers.
They combined visual object detection techniques with medical ontology--the Unified Medical Language System (UMLS)--to reduce the amount of manual labeling required to detect each disease, according to their paper published at the Journal of the American Informatics Association.
They mapped 2,220 images from Google search results depicting 32 diseases to the part of the body shown: eyes, ears, lips, hands or feet.
A previous method, supervised image retrieval, requires training the system with images of each disease individually. The UMLS lists 428 diseases of the eyes, the authors note, making training the system on each one too burdensome. By training the system to identify body parts, images can be reused across diseases.
They tested their method on two kinds of image sets: images of multiple diseases located on the same body part, and images of diseases that are located on more than one body part. Their method showed comparable precision to that of the supervised method while reducing the required manual labeling to just 15.6 percent. They improved overall precision 3.9 percent to 81.6 percent, and improved precision by 10 percent for 40.6 percent of the diseases.
The method has its limits, though. They carefully controlled the size and resolution of the images. And the system has trouble detecting skin, muscle, and veins, they said, making it difficult to expand the system to include those diseases. Next, they plan to build a two-tier system to fine-tune their approach.
The market for enterprise image archiving is expected to double by 2018, according to a report from Frost & Sullivan. It says the overall storage and archiving volume requirements for U.S. medical imaging data will surpass the one exabyte mark in 2016.
Maine's health information exchange HealthInfoNet has been a leader in creating a statewide medical imaging archive, partnering with Dell to store images in a public cloud.
To learn more:
- find the abstract