Automatic myeloblast segmentation in acute myeloid leukemia images based on adversarial feature learning.

• Presented a novel end-to-end segmentation architecture AMLcGAN based on conditional generative adversarial net (cGAN) to segment acute myeloid leukemia (AML) in Wright-Giemsa-stained bone marrow aspirate smear images. • Compared with other three popular deep learning-based segmentation models, AMLcGAN can achieve the best segmentation performance. • We validated AMLcGAN on a specific diagnosis task, the comparative clinical indicators show that presented model can provide accurate assistance for pathologists in the diagnosis of AML. The diagnosis of acute myeloid leukemia (AML) requires reliable detection and counting of immature white blood cells called myeloblasts on bone marrow aspirate and biopsy slides. Experienced hematologists/oncologists typically spend many hours manually reviewing blood cells and performing a microscopic morphological examination of myeloblasts on glass slides to determine this diagnosis. While manual counting of myeloblasts is considered the gold standard for an accurate diagnosis of AML, the process is time-consuming and difficult to standardize due to high intra- and inter-observer variability. Machine learning algorithms coupled with deep learning techniques represent potentially promising approaches for automated counting. Since myeloblasts possess very similar texture patterns to other hematopoietic cells (OHCs), accurately differentiating myeloblasts from other blood cells is a challenge for most segmentation models. We present a new conditional generative adversarial network (cGAN)-based segmentation model (AMLcGAN) to efficiently segment myoblasts from slides of patients with AML. Validation of the approach on 204 cytological images reveals a mean dice coefficient (mDICE) of 82.51 % with respect to manual segmentations. We also compared the new AMLcGAN method with three deep learning models: Unet, Segnet and Pix2pix, the corresponding mDICE values were 77.4 %, 78.2 %, and 79.53 %, respectively. We also employed the segmentation results from 4 different models to extract blast statistics feature, Haralick texture feature, Fractal dimension feature and myeloblast shape features to create 4 different prognostic models to predict risk of recurrence in AML patients. TheConcordance Index values for AMLcGAN and the three other models respectively were 0.63, 0.53, 0.58, and 0.56. The comprehensive evaluation results presented suggest that the AMLcGAN approach can enable accurate blast segmentation and for construction of accurate prognostic models for predicting risk of recurrence in AML patients. [ABSTRACT FROM AUTHOR]