Please click here to download this publication
The inspection of abnormal white blood cells in Leukemia via pathological visual inspection of a blood smear on a microscope slide is essential to the diagnosis of this cancer. AI image detection may one day help with more accurate and faster screening in this process that is entirely reliant on human pathologists currently. This recently published paper used AI image training models to learn what Leukemia blast white cells look like and tested its accuracy by asking it to create its own cells.
21 739 pre-annotated images of acute myeloid leukemia blast cells were scanned on our Motic Infinity 60 scanner. 17431 images were used for training with 4308 images used for evaluation.
An open source StyleSwin generative adversarial network AI model was used for the training of the images with a total of 500000 steps.
The AI-generated blast cells displayed the classic morphological features such as high nuclear-cytoplasmic ratio and less condensed nuclei as well as accurate surrounding red blood cells. However, some AI-generated cells showed coarse pixelation of the nuclei not seen in the original cells which may be due to resolution limitations or the model itself.
Overall, this model can help in the future for rare disorders with limited datasets if the accuracy of the algorithm can be improved. Using automated scanners such as the Motic Infinity 60 can help aid the generation of image databases which are crucial for AI algorithm training.
