Enhancement of 2D U-Net biomedical segmentation model with ImageNet pretrained CNNs
Keywords: 2D U-Net, Engineered Heart Tissue (EHT), ResNet-50, VGG-16, InceptionResNetV2
TL;DR: An approach to improve 2D U-Net biomedical segmentation model with CNNs pre-trained on ImageNet
Abstract: Segmentation is one of the most important steps in medical and biomedical studies. It allows retrieval of the object and removal of unnecessary background. In this work, we propose the enhancement of 2D U-Net with ImageNet-pre-trained models (ResNet-50, VGG-16 and InceptionResNetV2) used as backbones. Part of the layers of these models were transferred to the 2D U-Net encoder (one CNN layers per experiment). The modified segmentation model was trained, tested, and evaluated on the set of frames from spatial video showing a single Engineered Heart Tissue (EHT) cell. The data set consisted of 1240, 140, and 100 frames for training, testing and evaluation, respectively. Two metrics were calculated during evaluation - in the best case Dice reached 0.945705 whilst Intersection over Union (IoU) was equal to 0.897424.
Submission Number: 5
Loading