Background: Texture features are valuable clues for skilled technicians to differentiate peripheral blood (PB) white blood cells (WBCs). Some studies have tried to distinguish WBCs automatically by using texture analysis. However, no study so far has applied a gray level co-occurrence matrix (GLCM) to images of PB WBCs. Here, we developed a new GLCM method, called the CM-Tx method, for automatically distinguishing PB WBCs.
Methods: We used a total of 199 images of six different types of PB WBCs, taken from PB smears of 12 healthy volunteers, as objective standard images for the analysis. The six types were band form neutrophil, segmented form neutrophil, eosinophil, basophil, lymphocyte, and monocyte. Using in-house FORTRAN programs, three types of GLCM (R: distance (d) = 1, direction (θ) = 0°), (U: d = 1, θ = 270°) and (AE: d = 1, θ = 15° x q: q = 0, ..., 23), the mean intensity (MI) of each image and nine different texture distinction indexes (TDIs) for each GLCM were calculated. Then, a threshold value (TV) for distinguishing the type of PB WBC was selected from the dot plots of all TDIs and the MI.
Results: In total, we made 1,194 GLCMs. Using the selected TVs of the TDI, four sequential binary divisions could distinguish five types of PB WBCs. First, monocytes were distinguished (sensitivity 100%, specificity 100%, p < 0.0001) with the TV of the inverse difference moment of the GLCM (U). Then, segmented and band form neutrophils were distinguished from the remaining (100%, 99%, p < 0.0001) with the TV of the contrast of the GLCM (AE). Next, lymphocytes were distinguished (100%, 98%, p < 0.0001) with the TV of the entropy of the GLCM (AE). Finally, basophils were distinguished (82.4%, 100%, p < 0.0001) from eosinophils with the TV of the summed entropy of the GLCM (R). Band form neutrophils could not be distinguished from segmented form neutrophils. The average sensitivity of the CM-Tx method for the five types was 95.6%, and its average specificity was 99.3%.
Conclusions: The CM-Tx method can distinguish five types of PB WBCs by using numerical differences only in texture futures quantified with GLCM. However, some other method was needed to distinguish the band and segmented form neutrophils from each other.