Tackling the class imbalance problem of deep learning-based head and neck organ segmentation.

PURPOSE: The segmentation of organs at risk (OAR) is a required precondition for the cancer treatment with image- guided radiation therapy. The automation of the segmentation task is therefore of high clinical relevance. Deep learning (DL)-based medical image segmentation is currently the most successful approach, but suffers from the over-presence of the background class and the anatomically given organ size difference, which is most severe in the head and neck (HAN) area. METHODS: To tackle the HAN area-specific class imbalance problem, we first optimize the patch size of the currently best performing general-purpose segmentation framework, the nnU-Net, based on the introduced class imbalance measurement, and second introduce the class adaptive Dice loss to further compensate for the highly imbalanced setting. RESULTS: Both the patch size and the loss function are parameters with direct influence on the class imbalance, and their optimization leads to a 3% increase in the Dice score and 22% reduction in the 95% Hausdorff distance compared to the baseline, finally reaching [Formula: see text] and [Formula: see text] mm for the segmentation of seven HAN organs using a single and simple neural network. CONCLUSION: The patch size optimization and the class adaptive Dice loss are both simply integrable in current DL-based segmentation approaches and allow to increase the performance for class imbalance segmentation tasks.

Training of head and neck segmentation networks with shape prior on small datasets.

PURPOSE: Cancer in the head and neck area is commonly treated with radiotherapy. A key step for low-risk treatment is the accurate delineation of organs at risk in the planning imagery. The success of deep learning in image segmentation led to automated algorithms achieving human expert performance on certain datasets. However, such algorithms require large datasets for training and fail to segment previously unseen pathologies, where human experts still succeed. As pathologies are rare and large datasets costly to generate, we investigate the effect of: reduced training data, batch sizes and incorporation of prior knowledge. METHODS: The small data problem is studied by training a full-volume segmentation network with the reduced amount of data from the MICCAI 2015 head and neck segmentation challenge. To improve the segmentation, we evaluate the batch size as a hyper-parameter and first study and then incorporate a stacked autoencoder as shape prior into the training process. RESULTS: We found that using half of the training data (12 images of 25) results in an accuracy drop of only 3% for the segmentation of organs at risk. Also, the batch size turns out to be relevant for the quality of the segmentation when trained with less than half of the data. By applying PCA on the autoencoder's latent space we achieve a compact and accurate shape model, which is used as a regularizer and significantly improves the segmentation results. CONCLUSION: Small training data of up to 12 training images is enough to train accurate head and neck segmentation models. By using a shape prior for regularization, the performance of the segmentation can be improved significantly on the full dataset. When training on fewer than 12 images, the batch size is relevant and models have to be trained much longer until convergence.

3-D graph cut segmentation with Riemannian metrics to avoid the shrinking problem.

Though graph cut based segmentation is a widely-used technique, it is known that segmentation of a thin, elongated structure is challenging due to the "shrinking problem". On the other hand, many segmentation targets in medical image analysis have such thin structures. Therefore, the conventional graph cut method is not suitable to be applied to them. In this study, we developed a graph cut segmentation method with novel Riemannian metrics. The Riemannian metrics are determined from the given "initial contour," so that any level-set surface of the distance transformation of the contour has the same surface area in the Riemannian space. This will ensure that any shape similar to the initial contour will not be affected by the shrinking problem. The method was evaluated with clinical CT datasets and showed a fair result in segmenting vertebral bones.