Abstract:
This study targets the diagnosis of three features of a bone marrow report in hopes that in future, a
complete bone marrow diagnosis can be given through artificial intelligence. The three features being
identified are cellularity, megakaryocyte count and reticulin presence.
Currently, the diagnosis of a bone marrow trephine is given by a highly specialised doctor which
are known by the title of haematologist in the medical field. These specialists have years of experience and
study under their belt and even then, there is intra-observer and interobserver variability in the diagnosis.
The reason behind that is the diagnosis is observational and no hard rule exists for determining most of the
features of bone marrow trephine, and for those where a hard rule exists, the results are still given by
observation. Observational results will always differ from doctor to doctor and time to time for the same
doctor as well. This research aims to minimize interobserver and interobserver variability in bone marrow
trephine diagnosis.
The data for this study was collected manually which was a time taking part and consumed a major
chunk of the time dedicated to this study. The data was collected from Armed Forces Bone Marrow
Transplant Centre (AFBMTC), Combined Military Hospital (CMH), Rawalpindi, Pakistan. The data
collected was then cleaned and labelled and used for training different algorithms.
Another aim of this study was to calculate all three features from a single algorithm, for this purpose two
algorithms were used and compared, one based on intensity and texture features and the other on ResNet50
feature vectors. The first method outperformed the other for cellularity. The second method outperformed
the first on reticulin while neither of them performed good for megakaryocytes.
