Segmentation of White Blood Cells in M1, M2, and M3 Acute Myeloid Leukemia (AML) Images Using Deep Convolutional Neural Network

Leukemia is a cancer of the spinal cord and blood. A haematology analyzer diagnoses leukemia based on manual counting of white blood cells (WBCs) so it takes a lot of time and costs. Several methods with mathematical basis have been developed for detecting undeveloped cells so that can provide good performance for the segmentation. This research aims to segment WBCs on M1, M2, and M3 Acute Myeloid Leukemia (AML) images and test the performance results of segmentation using the Deep Convolutional Neural Network (Deep CNN) with previous studies used Active Contour without Edge (ACWE) and Watershed Distance Transform (WDT) for segmentation. The Deep CNN used was DeepLab v3+ with pretrained ResNet-50. M1, M2, and M3 AML images were taken from the Dr. Sardjito Hospital laboratory in Yogyakarta. The Fiji application was used to create ground truth mask images from the original data as a true label so they could be trained in the DeepLab v3+ model with pretrained ResNet-50 in 5-fold proportions of data with K-Fold Cross Validation and used the different number of filters. The results of the training model were used for new data segmentation of M1, M2, and M3 AML images. The highest average accuracy and average mIoU 5-fold cross-validation results between filters 64, 128, and 256 were obtained by the number of filters 256 of 0.9101 (91.01%) and 0.9060 (90.60%). This result surpassed previous research which used ACWE and WDT methods with an increase of 0.0722 (7.22%) for the number of filters 256.