Analysis of Peripheral Blood Films
Background
Peripheral blood film (PBF) remains a cornerstone of hematological diagnostics, enabling the identification of conditions ranging from anaemia to acute leukaemia. Traditionally, the standard practice in haematology laboratories has been to prepare PBFs on glass slides and perform light microscopy review by laboratory professionals or haematologists. While manual microscopy (MM) is the gold standard, it is labour-intensive and subject to inter-observer variability.
By integrating a digital microscope into the automated line, the glass slides can be scanned and stored as digital images, which facilitates quality assurance, education, image analysis, research, and archiving. As a next step, the digital images of blood cells can undergo automated analysis by machine learning models, which enables more rapid and accurate assessment in the hematology laboratory setting.
Approach
The computer vision AI model ("Blade") utilizes a deep convolutional neural network (CNN) for features extraction and object detection. To train the detector, regions of size 256x256 were cropped from the labelled PBFs. Grayworld normalization was applied to standardize the colour tone variations across different scanners. Pseudo labelling, a native semi-supervised learning technique, was used to leverage both labelled and unlabelled data to improve model performance. Blade was developed with 14 categories: Neutrophil, Lymphocyte, Large Granular Lymphocyte, Reactive Lymphocyte, Monocyte, Eosinophil, Basophil, Metamyelocyte, Myelocyte, Blast, Erythroblast, Smudge Cell, Artifact, and Giant Platelet.
For the evaluation, PBFs were randomly selected from routine laboratory bench samples from the haematology laboratory at Tan Tock Seng Hospital, Singapore. Differential counts were obtained by assessing 200 WBCs per slide. Two lab-certified medical technologists reviewed each slide using manual microscopy (MM). The same slides were then assessed by Blade, with excellent classification accuracy across most cell types when compared against the gold standard.
Demo