Blood Segmentation in Crossmatch Test Images Using Deep Learning

In blood transfusion services, crossmatch testing (pretransfusion) is essential to prevent patient complications. However, the main problem faced is the reading and documentation of the crossmatched test results which are still done manually. This can hamper hospital services if not supported by an adequate number of health workers. Therefore, this research aims to help hospital services so that recording medical records is no longer manual. It can allocate health workers more efficiently through the automation of reading and storing the results of the crossmatch test with the deep learning model of image segmentation that has been made. The methods used in the research include data collection, data preprocessing and augmentation, model training, and model evaluation. The model used in this research implements an encoder and decoder-type architecture. In the experiment, three types of encoder and decoder are implemented, such ResNet34, ResNeXt50, and EfficientNet for the encoder and Feature Pyramid Network (FPN), Pyramid Attention Network (PAN), and DeepLabV3+ for the decoder. The result shows that the best model used ResNeXt50 encoder and PAN decoder with IoU score, F1 score, precision, recall, and accuracy of 0.9371, 0.8636, 0.9152, 0.9238, 0.9811, respectively.