Therapeutic measures are guided by percentages of tumour blasts in bone marrow smears in patients with Acute Lymphoblastic Leukaemia (ALL). Recognising these blasts from benign blasts requires experience and time. Furthermore the inter-observer variability remains large. If a neural network could be trained to identify and count tumour blasts, this would lead to increased efficiency, savings in time and inter-observer variability might be reduced.
Prof. Delabie collected 30 whole slide scans of bone marrow smears. All patient data was omitted. On these whole slide scans cells were annotated by Prof. Delabie and his team with the IxorThink platform.
IxorThink then developed and trained an object detection model based upon the Detectron object detection model. This detection model extracts cells of interest from the whole slide image. Then a convolutional neural network extracts features of all detected cells. Finally a XBGBoost classifier is trained to classify each detected cell as blast, non-blast or unidentifiable cell.
The results of the model will be validated upon new whole slide images. When this validation is positive, both partners have the intention to work together on a project, which will aim to realise an end-to-end solution using machine learning to define and count tumour blasts in digital images of bone marrow smears with accuracy close to human accuracy level.