This paper presents a novel approach that automatically performs the immediate segmentation and classification of red blood cell storage lesions according to the storage period from phase images obtained with a digital holographic microscope. As a new approach, Pix2Pix-based network among deep learning technologies was proposed, which was compared with the existing U-net network. Our method showed excellent performance in red blood cell multi-class segmentation and classification accuracy with a high throughput of about 152 cells per second and a Dice coefficient of about 0.94. In addition, to evaluate the trained model, our method was applied to red blood cell images of 11 storage periods. In red blood cell images according to storage period, our method not only showed an excellent performance of about 95% in the confusion matrix, but was also in good agreement with previous results regarding changes in red blood cell markers(dominant shape) with storage period. Therefore, our method is believed to be useful as an automatic segmentation and classification method of red blood cell storage lesions by showing excellent performance in the identification of red blood cell storage lesions for safe blood transfusion.
Table Of Contents
Ⅰ. INTRODUCTION 1 Ⅱ. DIGITAL HOLOGRAPHIC MICROSCOPY(DHM) AND SAMPLE PREPARATION 6 2.1 Label-free off-axis DHM 6 2.2 Sample preparation 9 Ⅲ. DATA PREPROCESSING 10 Ⅳ. PROPOSED DEEP LEARNING MODEL 12 4.1 U-net 12 4.2 Pix2Pix 13 4.3 Construction of the patches and augmentation 16 4.4 Marker-based watershed algorithm 16 Ⅴ. EVALUATION METRIC 17 Ⅵ. EXPERIMENT RESULTS 18 6.1 Implementation details 18 6.2 Evaluation of multi-class RBC segmentation and classification performance 20 6.3 Classifications of RBCs based on storage durations & automated analysis of RBCs aging markers 20 Ⅶ. CONCLUSION 23