Early Experience in Developing a Machine-Learning and Digital Pathology Approach to Automate Bone Marrow Differential Counts
Year: 2018
Type: article
Abstract: Manual bone marrow aspirate (BMA) differential cell counts (DCCs) are often critical for diagnosis and classification of hematologic disorders, yet time-consuming and partly subjective with inter-/intraobserver and sampling variability. Attempts to automate this activity with current technologies, such as blood hematology analyzers, have thus far proven unsuccessful. Digital imaging technology coupled with machine learning algorithms for automated cell identification and counting represents a highly promising, emerging technology for this purpose. However, training data sets for BMA cellular constituents, critical for building and validating machine learning algorithms, are lacking and difficult to obtain. Our objective is to create a training data set of the cell types encountered on BMA smears from which algorithms can be trained to detect and classify different cell types. Whole slide images of Wright-stained BMA smears (×400) were obtained using a slide scanner. Images were uploaded to HistomicsTK, a web-based system for the management and annotation of digital pathology images that we developed. Five to 10 randomly selected nonoverlapping regions of interest (ROIs) from 12 BMA smears were annotated by a single experienced observer, thus localizing and assigning each cell a respective class. In total, 8,787 annotations collected from 76 ROIs comprised 12 cell classes: erythroid precursors (1,393), blasts (286), promyelocytes (108), myelocytes (411), metamyelocytes (441), bands/neutrophils (993), eosinophils (154), basophils (21), monocytes (109), lymphocytes (362), plasma cells (83), and megakaryocytes (14). Poorly preserved, unclassifiable cells and naked/smudged nuclei were annotated as unknown (3,160). Areas with red blood cells and debris within ROIs were carefully circled for exclusion. Challenges we discovered include exhaustively annotating every cell in the ROIs, to avoid bias, which yields higher numbers of cells in the unknown category, and excluding often abundant red blood cells and debris that were intermingled with the nucleated cells of interest. The other main challenges for automating cell counts are detecting and localizing cells with complex and varying morphologies, distinguishing closely packed cells, and employing localizations for cell classification. The large number of classes, subtle interclass differences, and variations in smear preparation and staining (color and contrast) represented additional, distinct obstacles we identified for BMA analysis. Lastly, the collected data set had an imbalanced distribution among classes, which was unavoidable due to the natural uneven distribution of cell types. This creates a problem in training the algorithm, as only a few features are learned from classes that have limited data. We addressed this problem by annotating additional cells from these classes, without the constraint of ROIs, and will use oversampling and augmentation techniques when training the algorithms. Having created a promising training data set, we are testing this exciting new algorithm on another data set with the ultimate goal of devising a reliable, objective method that automates DCCs.
Institution: Emory University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W2892035680