Constantin Ulrich

TL;DR: The pursuit of novel architectures in 3D medical image segmentation often overshadows the importance of rigorous validation. Many recent claims fail to hold up when scrutinized for common validation shortcomings. The study finds that employing CNN-based U-Net models, using the nnU-Net framework, and scaling models to modern hardware resources achieve state-of-the-art performance.

TL;DR: Transformer Ablations as mentioned in this paper replace the Transformer blocks with plain linear operators to quantify the effectiveness of Transformers in medical image segmentation, and explore the replaceable nature of Transformer-learnt representations.

TL;DR: This paper proposed to promote dissimilarity at different depths between architectures, with the goal of learning robust ensembles with disjoint failure modes, and showed that highly dissimilar intermediate representations result in less correlated output predictions and slightly lower error consistency, resulting in higher ensemble accuracy.