Learning Agreement from Multi-source Annotations for Medical Image Segmentation

In medical image analysis, it is typical to merge multiple independent annotations as ground truth to mitigate the bias caused by individual annotation preference. However, arbitrating the final annotation is not always effective because new biases might be produced during the process, especially when there are significant variations among annotations. This paper proposes a novel Uncertainty-guided Multi-source Annotation Network (UMA-Net) to learn medical image segmentation directly from multiple annotations. UMA-Net consists of a UNet with two quality-specific predictors, an Annotation Uncertainty Estimation Module (AUEM) and a Quality Assessment Module (QAM). Specifically, AUEM estimates pixel-wise uncertainty maps of each annotation and encourages them to reach an agreement on reliable pixels/voxels. The uncertainty maps then guide the UNet to learn from the reliable pixels/voxels by weighting the segmentation loss. QAM grades the uncertainty maps into high-quality or low-quality groups based on assessment scores. The UNet is further implemented to contain a high-quality learning head (H-head) and a low-quality learning head (L-head). H-head purely learns with high-quality uncertainty maps to avoid error accumulation and keeps strong prediction ability, while L-head leverages the low-quality uncertainty maps to assist the backbone to learn maximum representation knowledge. UNet with H-head will be reserved during the inference stage, and the rest of the modules can be removed freely for computational efficiency. We conduct extensive experiments on an unsupervised 3D segmentation task and a supervised 2D segmentation task, respectively. The results show that our proposed UMA-Net outperforms state-of-the-art approaches, demonstrating its generality and effectiveness.

Rethinking annotation granularity for overcoming deep shortcut learning: A retrospective study on chest radiographs

Deep learning has demonstrated radiograph screening performances that are comparable or superior to radiologists. However, recent studies show that deep models for thoracic disease classification usually show degraded performance when applied to external data. Such phenomena can be categorized into shortcut learning, where the deep models learn unintended decision rules that can fit the identically distributed training and test set but fail to generalize to other distributions. A natural way to alleviate this defect is explicitly indicating the lesions and focusing the model on learning the intended features. In this paper, we conduct extensive retrospective experiments to compare a popular thoracic disease classification model, CheXNet, and a thoracic lesion detection model, CheXDet. We first showed that the two models achieved similar image-level classification performance on the internal test set with no significant differences under many scenarios. Meanwhile, we found incorporating external training data even led to performance degradation for CheXNet. Then, we compared the models' internal performance on the lesion localization task and showed that CheXDet achieved significantly better performance than CheXNet even when given 80% less training data. By further visualizing the models' decision-making regions, we revealed that CheXNet learned patterns other than the target lesions, demonstrating its shortcut learning defect. Moreover, CheXDet achieved significantly better external performance than CheXNet on both the image-level classification task and the lesion localization task. Our findings suggest improving annotation granularity for training deep learning systems as a promising way to elevate future deep learning-based diagnosis systems for clinical usage.