Test-Time Augmentation Towards Improved Cross-Center PET/CT Segmentations by the Swin UNETR

In this study, we performed lymphoma lesion segmentation and make a dual contribution. First, we evaluate the impact of Test-Time Augmentation (TTA) on segmentation performance of the Swin UNEt TRansformer (UNETR). Secondly, we assess the effectiveness of TTA on a model trained with data from two different centers and on different lymphoma subtypes. Our results show that TTA significantly improves the segmentation performance of the Swin UNETR model (with an average Dice of 0.70 on the test data of center I and II) from Dice=0.55±0.09 to 0.61±0.17 (p-value<0.001) on the external unseen data from center III. Moreover, we also showed that TTA improves the generalizability of the model, enabling it to perform well on data from different centers and lymphoma types. We found the weights of the transformation for TTA by random serach. These findings have important implications for improving the accuracy and robustness of lymphoma lesion segmentation in clinical practice.