For
some diseases, assessment of historical medical records in the database is
sufficient for quick and accurate diagnosis. Many of such historical medical records
are in form of images, such as images of affected body parts of the patients
indicating a disease or abnormality. The doctors can study new images by comparing
them to similar images available in the database. Moreover, since such
historical records are generally stored along with their corresponding
diagnoses in the database, it becomes easier for the doctors to diagnose a
patient.
As
the number of records in the database increases, the database may become
comprehensive and exhaustive resulting into a consequent improvement in
accuracy of a diagnosis based on the database. However, handling of such large
amount of data poses a challenge. It is difficult to implement an architecture
that enables archiving of such large number of records that allows quick
retrieval of relevant records on demand at low cost.
Xanadu
is a big data management technology. Xanadu provides resilient, durable, scalable,
and consistent distributed big data database. Xanadu enables competitive big
data management in the clouds or enterprises.
Xanadu’s
high scalability makes it an ideal choice for the above mentioned type of
problem. Xanadu distributed hash means thousands of images can be stored and
retrieved efficiently using commodity hardware with a very low cost per GB.
Xanadu’s query system can also be leveraged to retrieve images quickly even
when the total images in the database grows into the millions, or even
billions. For details regarding Xanadu, please see following references.
A
content based image retrieval (CBIR) system enables search and retrieval of
images similar to a target image in large databases base on contents of images
(e.g. colors, shapes, textures). A common use-case of CBIR in medical diagnosis
is where imaging methods are used to highlight small areas (lesions) in
otherwise healthy tissue. Early breast cancer can be seen as small shadows on a
Mamogram (X-Ray of the breast), PET scans highlight small areas of increased
metabolic activity that can characterize cancerous growths and Retinal images
show small bleeds (microaneurysms) that highlight eye disease as well as wider
metabolic problems such as type II diabetes.
To
demonstrate the performance and capability of Xanadu based medical big data CBIR
system, a prototype CBIR system for retinal images is developed. In the case of
the retina (the light sensing surface at the rear of the eye), a simple
non-invasive photograph of the eye is sufficient to determine whether a patient
suffers from a range of diseases. Indeed, the signs of other diseases such as
diabetes, high blood pressure and other circulatory disorders can be diagnosed
and assessed from a single retinal image.
The
prototype CBIR system collects each retinal image together with its expert
reviewed diagnosis. Then, the system breaks each image into small patches that
are as small as possible without losing the ability to contain the typical
lesions that can indicate disease. The system utilizes the PCA technique to
cluster images together in a way that naturally groups similar images. For
improving the accuracy, the system uses machine learning techniques (e.g.
random forest or deep neural networks). The machine learning techniques also enable
to infer an overall diagnosis of a patch given the disease “score” and
“distance” (in image pixel terms) from known examples. The result is a detailed
(pixel by pixel) report for doctors where all areas of concern have been
highlighted and a detailed list of comparative images (showing the
similarities) can be viewed in the graphic user interface for final clinical
assessment.
If you are interested
in collaboration regarding medical big data CBIR applications utilizing medical
big data archive, please let me know: Alex G. Lee (alexglee@xanadubigdata.com)
Posts
No comments:
Post a Comment