Learning with limited annotations for medical image classification – MIMIC

Nowadays, Deep Learning has achieved a breakthrough in the field of
Artificial Intelligence. However, huge, labeled datasets are needed to
train on. Collecting such extensive annotated data is time and resources
consuming and it is not feasible in the medical domain. Indeed, unlike
the case of natural images, where annotations can be easily performed by
non-experts, medical images require careful and time-consuming analysis
from experts such us radiologists. The limited availability of annotated
medical imaging data remains the biggest challenge for the success of
deep learning techniques in medical imaging (i.e., “little (annotated)
data” challenge). Based on the idea that human can learn from few
samples, MIMIC project aims to propose low-data deep learning approaches
for few -shot volumetric medical images (such as MRI and CT)
classification. In MIMIC, we will tackle, as application, the problem of low prevalence
brain disorders and new emerging pneumonia detection as use cases for
the validation of developed approaches. However, we expect the general
proposed methodological development to be usable for other diseases
detection problems using only very small imaging data. The ultimate goal
of MIMIC is to help reducing medical data labeling effort for early and
efficient imaging-based disease diagnosis.