Neural network analysis of flow cytometry immunophenotype data

• Published in: IEEE transactions on biomedical engineering Vol. 43; no. 8; pp. 803 - 810
• Main Authors: Kothari, R., Cualing, H., Balachander, T.
• Format: Journal Article
• Language: English
• Published: New York, NY IEEE 01.08.1996; Institute of Electrical and Electronics Engineers
• Subjects: Adult; Algorithms; Biological and medical sciences; Biomedical measurements; Bone Marrow Cells; Cell Lineage; Cells (biology); Child; Computer science; Computerized, statistical medical data processing and models in biomedicine; Female; Flow Cytometry; Fluorescence; Fuzzy Logic; Hematology; Humans; Immune system; Immunophenotyping - methods; Investigative techniques, diagnostic techniques (general aspects); Leukemia - classification; Light scattering; Male; Medical diagnostic imaging; Medical management aid. Diagnosis aid; Medical sciences; Morphology; Neural networks; Neural Networks (Computer); Pathology. Cytology. Biochemistry. Spectrometry. Miscellaneous investigative techniques; Prognosis; Protocols; Antigen; Human; Flow cytometry; Phenotype; Leukemia; Classification; Exploration; Malignant hemopathy; Neural network; Cell differentiation; Immune system
• ISSN: 0018-9294
• DOI: 10.1109/10.508551

Acute leukemia is one of the leading malignancies in the United States with a mortality rate strongly influenced by the phenotype. This phenotype is based on detection of cell associated antigens normally expressed during leucopoietic differentiation. In this regard, leukemia classified as lymphoid or myeloid by phenotype is also classified as a candidate for the corresponding chemotherapy protocol. Additionally, the subtype of leukemia based on the degree of differentiation and cell maturity influence prognosis, response to treatment, and median survival times. In this paper, we analyze immunophenotype flow cytometry data toward categorization of leukemia into subcategories based on lineage and differentiation antigen expression. Twenty-eight inputs (derived from the mean fluorescence intensity of up to 27 antibodies, and an additional binary input denoting the past diagnosis of leukemia) are used as input to a neural classifier to categorize a total of 170 cases into the lineage and differentiation categories of leukemia. The neural classifier consisted of a feed forward network trained using back propagation. A complexity regulation term (weight decay) was used to improve the generalization performance of the neural classifier. A training error of 0.0% and a generalization error of 10.3% was obtained for categorization based on lineage, while a training error of 0.0% and a generalization error of 10.0% was obtained for categorization based on differentiation. These results indicate that objective classification of multifaceted phenotypes in leukemia can be achieved for analyzing multiparameter data in flow cytometry and further categorization into the prognostic subtypes.