Automatic Tumor Segmentation by Means of Deep Convolutional U-Net With Pre-Trained Encoder in PET Images

To assist physicists in developing radiation therapy treatment plans and in evaluating the effects of radiotherapy, an accurate and automatic tumor segmentation approach in positron emission tomography (PET) images is highly demanded in the clinical practice. In the present paper we investigate and construct a neural network architecture for auto-segmenting tumors by leveraging a 14-layer U-Net model with two blocks of a VGG19 encoder pre-trained with ImageNet. For pursuing efficient learning, a series of training strategies are proposed with limited training data. First, we apply a loss function based on Jaccard distance to re-balance the weights of training samples without re-weighting. Because of highly unbalanced data, re-weighting is an essential step but brings additional computation when cross-entropy loss function is used for medical image segmentation. Second, we import the DropBlock technique to replace the normal regularization dropout method as the former can help the U-Net efficiently avoid overfitting. We use a database containing 1309 PET images to train and test the proposed model. The mask, contour, and smoothed contour of a tumor are used as truths for teaching the proposed model. These are provided by expert radiologists. The segmentation accuracy compared to the truths is evaluated by calculating the Dice coefficient, Hausdorff distance, Jaccard index, sensitivity, and precision metrics. Extensive experimental results show that our method has achieved a relatively competitive performance in PET images on tumor segmentation. The volumes of the segmented tumors provided by our model would enable accurate automated identification and serial measurement of tumor volumes in PET images.