Gut Daniel, Tabor Zbislaw, Szymkowski Mateusz, Rozynek Milosz, Kucybala Iwona, Wojciechowski Wadim
IEEE Trans Med Imaging. 2022 Nov;41(11):3231-3241. doi: 10.1109/TMI.2022.3180435. Epub 2022 Oct 27.
In recent years, there were many suggestions regarding modifications of the well-known U-Net architecture in order to improve its performance. The central motivation of this work is to provide a fair comparison of U-Net and its five extensions using identical conditions to disentangle the influence of model architecture, model training, and parameter settings on the performance of a trained model. For this purpose each of these six segmentation architectures is trained on the same nine data sets. The data sets are selected to cover various imaging modalities (X-rays, computed tomography, magnetic resonance imaging), single- and multi-class segmentation problems, and single- and multi-modal inputs. During the training, it is ensured that the data preprocessing, data set split into training, validation, and testing subsets, optimizer, learning rate change strategy, architecture depth, loss function, supervision and inference are exactly the same for all the architectures compared. Performance is evaluated in terms of Dice coefficient, surface Dice coefficient, average surface distance, Hausdorff distance, training, and prediction time. The main contribution of this experimental study is demonstrating that the architecture variants do not improve the quality of inference related to the basic U-Net architecture while resource demand rises.
近年来，关于修改著名的U-Net架构以提高其性能有许多建议。这项工作的核心动机是在相同条件下对U-Net及其五个扩展进行公平比较，以厘清模型架构、模型训练和参数设置对训练模型性能的影响。为此，这六种分割架构中的每一种都在相同的九个数据集上进行训练。选择这些数据集是为了涵盖各种成像模态（X射线、计算机断层扫描、磁共振成像）、单类和多类分割问题以及单模态和多模态输入。在训练过程中，要确保所有比较的架构在数据预处理、数据集划分为训练集、验证集和测试子集、优化器、学习率变化策略、架构深度、损失函数、监督和推理等方面完全相同。性能根据骰子系数、表面骰子系数、平均表面距离、豪斯多夫距离、训练时间和预测时间进行评估。这项实验研究的主要贡献在于表明，架构变体在资源需求增加的同时，并没有提高与基本U-Net架构相关的推理质量。
