STAGE Bone Age Assessment - Gleamer internship

Bone age is a metric used to infer the biological maturity of children and adolescents. Although
correlated with the true age of an individual, abnormal discrepancies between true age and bone age
is a strong indicator of growth disorders. Bone age assessment (BAA) is thus a central part of the
diagnosis of several endocrine and metabolic disorders affecting children. As of today, most BAA rely
on either the TW2[1] or GP[2] methods, which are based on the comparison of a radio of a patient’s
hand to huge atlases of radios of “normally-developed bones. BAA is both tiresome and
time-consuming for radiologists, and the results often subjective and varying from a practitioner to
the other.
On the other hand, in the recent years, deep neural networks and in particular deep convolutional
neural networks have managed to reach exceptional performances on a wide variety of image-based
benchmarks, including medical imagery analysis tasks. In particular, some end-to-end solutions to the
BAA problem have appeared in the literature [4, 5, 6]. The system is first trained on a set of radios
annotated by radiologists who based their estimations on the atlases mentioned above. Once
trained, it is able to predict with high accuracy the bone age that a radiologist would have associated
to a given hand radiograph.
This internship would first focus on replicating the results of the aforementioned methods, and
improving the state of the art using several classical computer vision techniques (attention
mechanisms, multi-task learning, transfer learning) as well as ideas and suggestions from the intern.