Semantic Segmentation of Venous on Deep Vein Thrombosis (DVT) Case using UNet-ResNet

Deep Vein Thrombosis (DVT) is caused by an abnormal condition of blood clots in the network of blood vessels. No accurate profile data has been found on the number of common DVT cases in Indonesia. Several studies were conducted in several hospitals but with small sample sizes. In common cases, the diagnosis of DVT is made using Doppler Ultrasonography to monitor the condition of blood flow through the veins. This study uses the UNet-ResNet Deep Learning model to semantically segment the venous area on a 2D ultrasound image. The segmentation model is built from the pre-trained model UNet with the encoder ResNet-34. The dataset is taken from phantoms, a human body parts simulation tool. Ultrasound image acquisition on the Phantom will use Ultrasound Telemed SmartUs EXT-1M, which is directly connected to a PC. The segmentation model from the training process was evaluated with the Intersection-over-Union score (IoU) and Dice Loss. The result of the IoU evaluation on the standard UNet model resulted in an IoU score of 81.22% and an assessment of the dice loss of 0.1341. The UNet segmentation model assessment results with the ResNet-34 encoder using the IoU score of 84.50% and the dice loss matrix evaluation of 0.0857. The ResNet-34 model as an encoder in the UNet architecture can improve segmentation accuracy.