In many medical imaging and classical computer vision tasks, the Dice score and Jaccard index are used to evaluate the segmentation performance. Despite the existence and great empirical success of metric-sensitive losses, i.e. relaxations of these metrics such as soft Dice, soft Jaccard and Lovász-Softmax, many researchers still use per-pixel losses, such as (weighted) cross-entropy to train CNNs for segmentation. Therefore, the target metric is in many cases not directly optimized. We investigate from a theoretical perspective, the relation within the group of metric-sensitive loss functions and question the existence of an optimal weighting scheme for weighted cross-entropy to optimize the Dice score and Jaccard index at test time. We find that the Dice score and Jaccard index approximate each other relatively and absolutely, but we find no such approximation for a weighted Hamming similarity. For the Tversky loss, the approximation gets monotonically worse when deviating from the trivial weight setting where soft Tversky equals soft Dice. We verify these results empirically in an extensive validation on six medical segmentation tasks and can confirm that metric-sensitive losses are superior to cross-entropy based loss functions in case of evaluation with Dice Score or Jaccard Index. This further holds in a multi-class setting, and across different object sizes and foreground/background ratios. These results encourage a wider adoption of metric-sensitive loss functions for medical segmentation tasks where the performance measure of interest is the Dice score or Jaccard index.
1 April 2020 in International journal of legal medicine
Convolutional neural networks (CNNs) for brain tumor segmentation are generally developed using complete sets of magnetic resonance imaging (MRI) sequences for both training and inference. As such, these algorithms are not trained for realistic, clinical scenarios where parts of the MRI sequences which were used for training, are missing during inference. To increase clinical applicability, we proposed a cross-modal distillation approach to leverage the availability of multi-sequence MRI data for training and generate an enriched CNN model which uses only single-sequence MRI data for inference but outperforms a single-sequence CNN model. We assessed the performance of the proposed method for whole tumor and tumor core segmentation with multi-sequence MRI data available for training but only $T_{1}$ -weighted ( $T_{\text{1}w}$ ) sequence data available for inference, using BraTS 2018, and in-house datasets. Results showed that cross-modal distillation significantly improved the Dice score for both whole tumor and tumor core segmentation when only $T_{\text{1}w}$ sequence data were available for inference. For the evaluation using the in-house dataset, cross-modal distillation achieved an average Dice score of 79.04% and 69.39% for whole tumor and tumor core segmentation, respectively, while a single-sequence U-Net model using $T_{\text{1}w}$ sequence data for both training and inference achieved an average Dice score of 73.60% and 62.62%, respectively. These findings confirmed cross-modal distillation as an effective method to increase the potential of single-sequence CNN models such that segmentation performance is less compromised by missing MRI sequences or having only one MRI sequence available for segmentation.