Classification of red blood cells as normal, sickle, or other abnormal, using a single image analysis feature

Sickle cell anemia is a disease for which there is currently no effective treatment. One method of evaluating clinical status is the counting of cell types based on morphology. There is a need for a rapid, reproducible method, superior to human inspection, for classification of these cells. Quantitative digital-image analysis is being applied to this need. Blood from 24 patients with sickle cell anemia (SS) and SC disease and ten hematologically normal volunteers (AA) was stressed by bubbling with nitrogen. One hundred fifty cells were analyzed from each sickle specimen, and 100 were analyzed from each nonsickle specimen. Expert observers classified each cell as normal (N), sickle (S), or other abnormal (A). Cells were analyzed with a custom, high-resolution image-analysis instrument. A total of 42 features including metric, optical density-derived, and textural features were extracted. The metric feature Form Factor (4 pi Area/Perimeter2) was selected by recursive partitioning analysis as the sole feature needed for segregating cells into the classes of N, A, and S. The agreement of automated classification (using cutpoints determined by recursive partitioning analysis) with a human expert for specimens from individuals with sickle cell anemia was 89% for N-, 73% for A-, and 92% for S-classified cells. For specimens from AA individuals, the agreement was 92% for N and 76% for A. For specimens from individuals with sickle cell anemia, rates of agreement between two human experts were compared and found to be 86% for N, 84% for A, and 80% for S. For specimens from AA individuals, the agreement was 90% for N and 87% for A.