Improving Biomedical Image AI Training and Analysis | AI News | Intel Software

Improving Biomedical Image AI Training and Analysis | AI News | Intel Software


I’m David Shaw, and
this is AI News. We’re no strangers to
the concept of using AI to assist in medical treatment. One of the challenges
is the data used for analyzing biomedical
images can be imbalanced, hindering effectiveness. What do we mean by imbalanced? Well, think of medical
imaging datasets. They’re mostly composed
of normal samples. For example, a
healthy set of lungs. And there’s only a small
percentage of abnormal ones, say if there’s a
lung disease present. Because most classifiers
focus on learning the largest classes, it can affect
training of machine learning classifiers
in a negative way. This leads to poor
classification accuracy and diagnosis. The good news is that
there are techniques for approaching this challenge. Subhashis Banerjee, an
Intel student ambassador, is conducting research into
this dataset challenge. It’s leading to some
promising solutions. His research focuses on
biomedical image analysis. It includes computer-aided
disease localization and segmentation. The objective was to address
this class imbalance. In some cases, a class
might be underrepresented. The methodology he adopted
explored the impact of the class imbalance on the
performance of [INAUDIBLE].. The main medical image
analysis problems include disease or
abnormality detection, region of interest segmentation,
and disease classification from real medical
image datasets. For this research
project, Subhashis is using Intel AI DevCloud
with Intel-optimized versions of Python, TensorFlow, and
Karos for implementation and training the models. The research conducted
includes two new approaches for dealing with
this imbalance issue. One, class learning. Using this approach,
the training process focuses on a single
classifier for each class. And two, face training
with pre-training on randomly oversampled
and undersampled datasets. Continuous rounds
of testing will determine the effects
of class imbalance on the classification
performance and make it possible to compare
other methods to discover which approach achieves the
best result. Check out the story in the links to
learn more about class balance. What do you think could
be the next solution? Thanks for watching, and
I’ll see you next week. [MUSIC PLAYING]

One thought to “Improving Biomedical Image AI Training and Analysis | AI News | Intel Software”

Leave a Reply

Your email address will not be published. Required fields are marked *