Mitigating False Predictions In Unreasonable Body Regions

TL;DR

The paper addresses poor generalization of 3D CT segmentation when test data have larger Field-of-View (FOV) than training data, which leads to anatomically implausible predictions. It introduces Region Loss, a training-time penalty guided by a Body Part Regression (BPR) model that maps axial slices to standardized positions, enabling suppression of predictions in invalid body regions. The method is validated in both multi-dataset (MD) and single-dataset (SD) training regimes, within the MultiTalent/nnU-Net framework, and shows reduced false positives in spurious regions (up to 85.7% in MD) along with improved Dice scores. This approach enhances cross-FOV generalization for CT organ/tumor segmentation and offers a robust alternative to postprocessing, with potential extensions to other 3D modalities and detection tasks.

Abstract

Despite considerable strides in developing deep learning models for 3D medical image segmentation, the challenge of effectively generalizing across diverse image distributions persists. While domain generalization is acknowledged as vital for robust application in clinical settings, the challenges stemming from training with a limited Field of View (FOV) remain unaddressed. This limitation leads to false predictions when applied to body regions beyond the FOV of the training data. In response to this problem, we propose a novel loss function that penalizes predictions in implausible body regions, applicable in both single-dataset and multi-dataset training schemes. It is realized with a Body Part Regression model that generates axial slice positional scores. Through comprehensive evaluation using a test set featuring varying FOVs, our approach demonstrates remarkable improvements in generalization capabilities. It effectively mitigates false positive tumor predictions up to 85% and significantly enhances overall segmentation performance.

Figures (4)

• Figure 1: a) shows an example of a training dataset with a limited Fiel of View (FOV) caused by the desire to reduce radiation dose, scanning time and costs. Typically, the final model is then applied to images from different sources and studies with varying FOVs and tends to make anatomical implausible predictions in unseen body parts. b) shows the FOV of the datasets of the Medical Decathlon as well as the distribution of the location of the upper and lower position of the target structure, which was determined using a Body Part Regression model.
• Figure 2: The proposed Region Loss achieves always the best (red) or second best (blue) mean Dice. Additionally, the Region Loss improves the boundaries of the lower two boxplot quartiles compared to default MultiTalent or default nnU-Net. This indicates, that the Region Loss succesfully improves inferior cases.
• Figure 3: The proposed Region Loss mitigates most anatomical implausible False Positive tumor predictions in the MD setting.
• Figure 4: a) An example of unreasonable hepatic liver tumor predictions of MultiTalent and nnU-Net in the head that our proposed Region Loss completely mitigates. b) Example spleen predictions for which MultiTalent is more robust compared to nnU-Net.