In order to verify the segmentation accuracy of ALL microimages through different loss functions trained by U-Net,accuracy, IOU score and F1 score were used to evaluatetraining, validation and test process. Three evaluation indexescould obtain an objective evaluation (Table 2). CDWN performed as the best on accuracy and IOU score evaluationduring the train and validation process. Although it was notthe most accurate one on the F1 score, a disparity less than0.01 was observed, compared with the highest accuracy. Fortest results, CDWN always performed as the best.The results in Table 2 show that creating the third class corresponding to the cell border, and enhancing U-Net learningcell border features through the weight map combined classweights and distance transformation weights, were beneficial to improve WBC segmentation result in ALL microscopicimages (Fig. 10)