Improving the Performance of Deep CNNs in Medical Image Segmentation with Limited Resources

Abstract

Convolutional neural networks (CNNs) have obtained enormous success in image segmentation, which is substantial in many clinical treatments. Even though CNNs have achieved state-of-the-art performances, most researches on semantic segmentation using the deep learning methods are in the field of computer vision, so the research on medical images is much less mature than that of natural images, especially, in the field of 3D image segmentation. Our experiments on CNN segmentation models demonstrated that, with modifications and tuning in network architecture and parameters, modified models would show better performances in the selected task, especially with limited training dataset and hardware. We have selected the 3D liver segmentation as our goal and presented a pathway to select a state-of-the-art CNN model and improve it for our specific task and data. Our modifications include the architecture, optimization algorithm, activation functions and the number of convolution filters. With the designed network, we used relatively less training data than other segmentation methods. The direct output of our network, with no further post-processing, resulted in the dice score of ~99 in training and ~95 in validation images in 3D liver segmentation, which is comparable to the state-of-the-art models that used more training images and post-processing. The proposed approach can be easily adapted to other medical image segmentation tasks.