• DMs for Segmentation 论文串烧

        Author: Sijin Yu

        DMs for Segmentation 论文串烧1. 用 Bernoulli 噪音的 DM 用于医学图像分割1.0 Abstract1.1 Model1.1.1 Problem Definition1.1.2 Framework of Diffusion Model1.1.3 Loss Function1.2 Experiment1.2.1 Dataset & Preprocessing1.2.2 消融实验1.2.3 横向对比2. 使用类别引导的 CDM 进行弱监督分割2.0 Abstract2.1 Motivation & Contribution2.1.1 Motivation2.1.2 Contribution2.2 Model2.2.1 Training Conditional Denoising Diffusion Model2.2.2 Gradient Map w.r.t Condition2.3 Experiment2.3.1 Dataset & Preprocessing2.3.2 横向对比2.3.3 消融实验3. 使用 DM 的血液动力学对比增强 MRI 生成的 feature map 引导分割3.0 Abstract3.1 Model3.2 Experiment3.2.1 Dataset & Preprocessing3.2.2 横向对比3.2.3 消融实验4. 魔改 DM 中 U-Net 的输入和结构的医学图像分割4.0 Abstract4.1 Model4.1.1 Cross Attention4.1.2 Multi-sized Transformer U-Net4.2 Experiment4.2.1 Dataset4.2.1 消融实验4.2.2 横向对比5. 使用 DM 做数据增强以提高分类性能5.1 Abstract5.2 Motivation & Contribution5.2.1 Motivation5.2.2 Contribution5.3 Model5.3.1 Unconditional Nuclei Structure Synthesis5.3.2 Conditional Histopathology Image Synthesis5.4 Experiment5.4.1 Dataset5.4.2 实验6. 使用 DM 中的 U-Net 的一层 embedding 用作分割和分类6.1 Abstract6.2 Motivation & Contribution6.2.1 Motivation6.2.2 Contribution6.3 Model6.4 Experiment6.4.1 Dataset6.4.2 横向对比6.4.3 探索实验7. 使用 DM 生成 bounding box, 然后用于 mask 的生成7.1 Abstract7.2 Motivation & Contribution7.2.1 Motivation7.2.2 Contribution7.3 Model7.4 Experiment8. 使用 DM 做分割, 并用 DS 证据理论来促进多模态融合8.1 Abstract8.2 Motivation & Contribution8.2.1 Motivation8.2.2 Contribution8.3 Model8.3.1 Parallel DDPM Path8.3.2 EIL & CDO8.4 Experiment8.4.1 Dataset8.4.2 对比实验9. 在同一个 DM 分割模型中多次采样, 融合每次的结果, 并衡量不确定性9.1 Abstract9.2 Motivation & Contribution9.2.1 Motivation9.2.2 Contribution9.3 Model9.3.1 Evolution Uncertainty9.4 Experiment9.4.1 Dataset9.4.2 消融实验和对比实验10. 开山鼻祖 MedSegDiff: 使用动态条件编码和可学习滤波器的 DM 医学图像分割10.1 Abstract10.2 Motivation & Contribution10.2.1 Motivation10.2.2 Contribution10.3 Model10.3.1 FF-Parser10.4 Experiment10.4.1 Dataset10.4.2 对比实验10.4.3 消融实验11. MedSegDiff-V2: 同时在空间域和频率域上做 Attn 的条件和嵌入融合, 使用 DM 直接生成分割11.1 Abstract11.2 Motivation & Contribution11.2.1 Motivation11.2.2 Contribution11.3 Model11.4 Experiment11.4.1 Dataset11.4.2 消融实验11.4.3 对比实验12. 使用对抗学习的自监督分割 (含有 DM)12.1 Abstract12.2 Motivation & Contribution12.2.1 Motivation12.2.2 Contribution12.3 Model12.3.1 Generation Module12.3.2 Loss Function12.4 Experiment12.4.1 Dataset12.4.2 对比实验12.4.3 消融实验13. 自然图像上的 DM 分割13.1 Abstract13.2 Motivation & Contribution13.2.1 Motivation13.2.2 Contribution13.3 Model13.4 Experiment13.4.1 Dataset13.4.2 消融实验13.4.3 对比实验Reference List

        1. 用 Bernoulli 噪音的 DM 用于医学图像分割

        [1] Tao Chen, Chenhui Wang, Hongming Shan. BerDiff: Conditional Bernoulli Diffusion Model for Medical Image Segmentation. MICCAI, 2023.

        1.0 Abstract

        1.1 Model

        1

        1.1.1 Problem Definition

        1.1.2 Framework of Diffusion Model

        2

        1.1.3 Loss Function

        最后,

        (13)Ltotal=LKL+λBCELBCE

        1.2 Experiment

        1.2.1 Dataset & Preprocessing

        1.2.2 消融实验

        3

        1.2.3 横向对比

        4


        2. 使用类别引导的 CDM 进行弱监督分割

        [5] Xinrong Hu, Yu-Jen Chen, Tsung-Yi Ho, Yiyu Shi. Conditional Diffusion Models for Weakly Supervised Medical Image Segmentation. MICCAI, 2023.

        2.0 Abstract

        2.1 Motivation & Contribution

        2.1.1 Motivation

        2.1.2 Contribution

        2.2 Model

        2.2.1 Training Conditional Denoising Diffusion Model

        2.2.2 Gradient Map w.r.t Condition

        通过 xt 复原 xt1 可以通过:

        (19)xt1(xt,t,y)=α¯t1(xt1α¯tϵ^(xt,y)α¯t)+1α¯t1ϵ^θ(xt,y)

        xt1(xt,t,y)y 的偏微分 xt1y 可以通过下式计算

        (20)xt1(xt,t,y)y|y=y1=limτ1xt1(xt,t,f(y1))xt1(xt,t,τf(y1)+(1τ)f(y0))1τ

        在实验中, 取 τ=0.95.

        2.3 Experiment

        2.3.1 Dataset & Preprocessing

        2.3.2 横向对比

        6

        表 1 是在 BraTS 上的横向对比, 表 2 是在 CHAOS 上的横向对比.

        2.3.3 消融实验

        7

        图为主要超参数的消融实验结果.


         

        3. 使用 DM 的血液动力学对比增强 MRI 生成的 feature map 引导分割

        [8] Tianxu Lv, Yuan Liu, Kai Miao, Lihua Li, Xiang Pan. Diffusion Kinetic Model for Breast Cancer Segmentation in Incomplete DCE-MRI. MICCAI, 2023.

        3.0 Abstract

        什么是 DCE-MRI, 以及什么是 pre-contrast images 和 post-contrast images?

        ChatGPT: 在动态对比增强磁共振成像(DCE-MRI)中,"pre-contrast images" 和 "post-contrast images" 指的是MRI扫描中的不同图像获取阶段,特别是在使用对比剂的情况下。

        1. 前对比图像(Pre-Contrast Images): 这些图像是在注射对比剂之前获取的。前对比图像用作基准,并通常用于解剖参考,以提供与对比剂注射后获取的图像进行比较的基准点。它们有助于评估对比增强前组织的特征。

        2. 后对比图像(Post-Contrast Images): 这些图像是在将对比剂注入患者的血液后获取的。对比剂突出显示了血管和各种组织内的血流分布。后对比图像显示了对比剂在组织内的摄取和分布方式,并有助于评估灌注、血管化和病变或异常的存在。这些图像用于跟踪组织特征的变化,评估对比增强的影响。

        DCE-MRI通常用于医学影像学,用于评估各种疾病,包括肿瘤特征化、血管化评估和组织灌注的评估。前对比和后对比图像在提供有关诊断和治疗规划的宝贵信息方面发挥着关键作用。

        8

        上图左边表示血液动力图响应函数和 Denoising Diffusion Process 之间的关系, 右边表示体素级别的强度随时间的变化.

        [⚠️注: 这里, x0,x1,,xk 是 DCE-MRI 扫描的图片的过程, x0xt 是扩散模型的前向过程, xtxt1xk 是扩散模型的反向过程, 即扩散模型的任务是: 给定前对比图像 x0, 生成后对比图像 xk.]

        3.1 Model

        9

        模型分为 Diffusion Module (图中 a) 和 Segmentation Module (图中 b).

        Diffusion Module 为经典的 DDPM, 以前对比图像 x0 到高斯噪音 xt 为前向过程, 以高斯噪音 xtx0 为反向过程. DM 被预训练后, DM 中间的隐藏输出 fdm 就包含了血液动力学响应函数的信息.

        Segmentation Module 由前四层 KineticBlock 和后四层 UpBlock 组成.

        KineticBlock 同时以 DM 的隐藏输出 fdm 和上一层 KineticBlock 的输出 fsm 为输入, 通过一个 Fusion Layer 融合它们:

        (21)f^=Fusion(fdm,fsm)=Concat(ReLU(BN(Wfdm));fsm)

        3.2 Experiment

        3.2.1 Dataset & Preprocessing

        3.2.2 横向对比

        10

        3.2.3 消融实验

        11

        fi 表示 DM 的第 i 阶段的 feature map.


         

        4. 魔改 DM 中 U-Net 的输入和结构的医学图像分割

        [11] G. Jignesh Chowdary, Zhaozheng Yin. Diffusion Transformer U-Net for Medical Image Segmentation. MICCAI, 2023.

        4.0 Abstract

        4.1 Model

        12

        4.1.1 Cross Attention

        13

        最后的输入被 reshape 成和 fM 一样的形状.

        4.1.2 Multi-sized Transformer U-Net

        14

        U-Net 的组成是 Multi-sized Transformer.

        input 先通过 Multi-sized window 的 Transformer, 一共有 K 条路, 所有路的加和进入 Shifted window, 得到 output.

        4.2 Experiment

        4.2.1 Dataset

        4.2.1 消融实验

        15

        4.2.2 横向对比

        16


         

        5. 使用 DM 做数据增强以提高分类性能

        [17] Xinyi Yu, Guanbin Li, Wei Lou, Siqi Liu, Xiang Wan, Yan Chen, and Haofeng Li. Diffusion-Based Data Augmentation for Nuclei Image Segmentation. MICCAI, 2023.

        5.1 Abstract

        5.2 Motivation & Contribution

        5.2.1 Motivation

        5.2.2 Contribution

        5.3 Model

        17

        生成模型由两个步骤组成:

        5.3.1 Unconditional Nuclei Structure Synthesis

        18

        Nuclei Structure 由 pixel-level semantic (像素级别语义)distance transform (距离变换) 两部分组成.

        因此, 一个 Nuclei Structure 是具有三个通道的, 和原始图像一样大的图像.

        5.3.2 Conditional Histopathology Image Synthesis

        19

        5.4 Experiment

        5.4.1 Dataset

        5.4.2 实验

        20


         

        6. 使用 DM 中的 U-Net 的一层 embedding 用作分割和分类

        [20] Héctor Carrión and Narges Norouzi. FEDD - Fair, Efficient, and Diverse Diffusion-Based Lesion Segmentation and Malignancy Classification. MICCAI, 2023.

        6.1 Abstract

        6.2 Motivation & Contribution

        6.2.1 Motivation

        6.2.2 Contribution

        6.3 Model

        21

        在 DM 中的 U-Net 中指定的一层获得 embedding, 它通过上采样以进行分割, 通过下采样以进行分类.

        6.4 Experiment

        6.4.1 Dataset

        6.4.2 横向对比

        22

        6.4.3 探索实验

        23


         

        7. 使用 DM 生成 bounding box, 然后用于 mask 的生成

        [22] Mengxue Sun, Wenhui Huang , and Yuanjie Zheng. Instance-Aware Diffusion Model for Gland Segmentation in Colon Histology Images. MICCAI, 2023.

        7.1 Abstract

        7.2 Motivation & Contribution

        7.2.1 Motivation

        7.2.2 Contribution

        7.3 Model

        24

        7.4 Experiment

        25

        26


         

        8. 使用 DM 做分割, 并用 DS 证据理论来促进多模态融合

        [23] Jianfeng Zhao and Shuo Li. Learning Reliability of Multi-modality Medical Images for Tumor Segmentation via Evidence-Identified Denoising Diffusion Probabilistic Models. MICCAI, 2023.

        8.1 Abstract

        8.2 Motivation & Contribution

        8.2.1 Motivation

        8.2.2 Contribution

        8.3 Model

        27

        EI-DDPM 模型由三个部分组成:

        8.3.1 Parallel DDPM Path

        28

        DDPM 用于生成分割图, 以某一模态的图像为条件.

        8.3.2 EIL & CDO

        使用下文中的方法:

        https://blog.csdn.net/yusijinfs/article/details/134427358

        将 T1, T2, Flair, T1ce 四种模态的分割结果做融合.

        8.4 Experiment

        8.4.1 Dataset

        8.4.2 对比实验

        29


         

        9. 在同一个 DM 分割模型中多次采样, 融合每次的结果, 并衡量不确定性

        [27] Jiacheng Wang, Jing Yang, Qichao Zhou, Liansheng Wang. Medical Boundary Diffusion Model for Skin Lesion Segmentation. MICCAI, 2023.

        9.1 Abstract

        9.2 Motivation & Contribution

        9.2.1 Motivation

        9.2.2 Contribution

        9.3 Model

        9.3.1 Evolution Uncertainty

        9.4 Experiment

        9.4.1 Dataset

        9.4.2 消融实验和对比实验

        30


         

        10. 开山鼻祖 MedSegDiff: 使用动态条件编码和可学习滤波器的 DM 医学图像分割

        [30] Junde Wu, Rao Fu, Huihui Fang, Yu Zhang, Yehui Yang, Haoyi Xiong, Huiying Liu, and Yanwu Xu. MedSegDiff: Medical Image Segmentation with Diffusion Probabilistic Model. MIDL, 2023.

        10.1 Abstract

        10.2 Motivation & Contribution

        10.2.1 Motivation

        10.2.2 Contribution

        10.3 Model

        31

        10.3.1 FF-Parser

        32

        (25)M=AM
        (26)m=F1[M]

        10.4 Experiment

        10.4.1 Dataset

        10.4.2 对比实验

        33

        10.4.3 消融实验

        34


         

        11. MedSegDiff-V2: 同时在空间域和频率域上做 Attn 的条件和嵌入融合, 使用 DM 直接生成分割

        [33] Junde Wu, Rao Fu, Huihui Fang, Yu Zhang, and Yanwu Xu. MedSegDiff-V2: Diffusion based Medical Image Segmentation with Transformer. arXiv preprint arXiv:2301.11798, 2023.

        11.1 Abstract

        11.2 Motivation & Contribution

        11.2.1 Motivation

        11.2.2 Contribution

        11.3 Model

        35

        11.4 Experiment

        11.4.1 Dataset

        11.4.2 消融实验

        36

        11.4.3 对比实验

        37


         

        12. 使用对抗学习的自监督分割 (含有 DM)

        [35] Boah Kim, Yujin Oh, Jong Chul Ye. Diffusion Adversarial Representation Learning for Self-supervised Vessel Segmentation. ICLR, 2023.

        12.1 Abstract

        12.2 Motivation & Contribution

        12.2.1 Motivation

        12.2.2 Contribution

        12.3 Model

        38

        一组数据由两个图像组成, x0a血管造影 (angiography), x0b背景 (background). 在数据采集时, 先采集 x0b, 然后给患者注射对比剂, 然后采集到 x0a, 在这个过程中可能由于患者的移动导致两张图不对齐.

        12.3.1 Generation Module

        生成模块由 N 个 ResnetBlock 组成. 每个 ResnetBlock 的计算是可切换的 (计算取决于当前是路径 A 还是 B). 令特征图为 vRB×C×H×W, B,C,H,W 分别为批量大小, 通道数, 高, 宽. 在可切换层的计算如下:

        最后, 模型的生成方式为:

        12.3.2 Loss Function

        39

        对于训练的描述如上图所示, 用到了三个损失函数 Ladv,Ldiff,Lcyc.

        最后, 总的损失有两个:

        12.4 Experiment

        12.4.1 Dataset

        12.4.2 对比实验

        40

        12.4.3 消融实验

        41


         

        13. 自然图像上的 DM 分割

        [36] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, Shalini De Mello.Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models. CVPR, 2023.

        13.1 Abstract

        13.2 Motivation & Contribution

        13.2.1 Motivation

        13.2.2 Contribution

        13.3 Model

        训练:

        42

        测试:

        43

        13.4 Experiment

        13.4.1 Dataset

        13.4.2 消融实验

        44

        13.4.3 对比实验

        45


         

         

         

         

         

         

         

         

         

         

         

         

         

         

         

         

         

         

         

         

         

         

         

        Reference List

        [1] Tao Chen, Chenhui Wang, Hongming Shan. BerDiff: Conditional Bernoulli Diffusion Model for Medical Image Segmentation. MICCAI, 2023.

        [2] Armato III, S.G., McLennan, G., Bidaut, L., McNitt-Gray, M.F., Meyer, C.R., Reeves, A.P., Zhao, B., Aberle, D.R., Henschke, C.I., Hoffman, E.A., et al. The lung image database consortium (LIDC) and image database resource initiative (IDRI): a completed reference database of lung nodules on CT scans. Medical physics, 2011.

        [3] Clark, K., Vendt, B., Smith, K., Freymann, J., Kirby, J., Koppel, P., Moore, S., Phillips, S., Maffitt, D., Pringle, M., et al. The cancer imaging archive (TCIA): maintaining and operating a public information repository. Journal of digital imaging, 2013.

        [4] Baid, U., Ghodasara, S., Mohan, S., Bilello, M., Calabrese, E., Colak, E., Fara- hani, K., Kalpathy-Cramer, J., Kitamura, F.C., Pati, S., et al. The RSNA-ASNR- MICCAI BraTS 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv:2107.02314, 2021.

        [5] Xinrong Hu, Yu-Jen Chen, Tsung-Yi Ho, Yiyu Shi. Conditional Diffusion Models for Weakly Supervised Medical Image Segmentation. MICCAI, 2023.

        [6] Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J.S., Freymann, J.B., Farahani, K., Davatzikos, C. Advancing the cancer genome atlas glioma mri collections with expert segmentation labels and radiomic features. Scientific data, 2017.

        [7] Kavur, A.E., Gezer, N.S., Barı ̧s, M., Aslan, S., Conze, P.H., Groza, V., Pham, D.D., Chatterjee, S., Ernst, P., O ̈zkan, S., Baydar, B., Lachinov, D., Han, S., Pauli, J., Isensee, F., Perkonigg, M., Sathish, R., Rajan, R., Sheet, D., Dovletov, G., Speck, O., Nu ̈rnberger, A., Maier-Hein, K.H., Bozdag ̆ı Akar, G., U ̈nal, G., Dicle, O., Selver, M.A. CHAOS Challenge - combined (CT-MR) healthy abdominal organ segmentation. Medical Image Analysis, 2021.

        [8] Tianxu Lv, Yuan Liu, Kai Miao, Lihua Li, Xiang Pan. Diffusion Kinetic Model for Breast Cancer Segmentation in Incomplete DCE-MRI. MICCAI, 2023.

        [9] Newitt, D., Hylton, N. Single site breast DCE-MRI data and segmentations from patients undergoing neoadjuvant chemotherapy. Cancer Imaging Arch, 2016.

        [10] Hyun-Jic Oh, Won-Ki Jeong. DiffMix: Diffusion Model-Based Data Synthesis for Nuclei Segmentation and Classification in Imbalanced Pathology Image Datasets. MICCAI, 2023.

        [11] G. Jignesh Chowdary, Zhaozheng Yin. Diffusion Transformer U-Net for Medical Image Segmentation. MICCAI, 2023.

        [12] Jha, D., et al. Kvasir-SEG: a segmented polyp dataset. Springer, Cham, 2020.

        [13] Bernal, J., S ́anchez, F.J., Fern ́andez-Esparrach, G., Gil, D., Rodr ́ıguez, C., Vilarin ̃o, F. Wm-dova maps for accurate polyp highlighting in colonoscopy: valida- tion vs. saliency maps from physicians. Comput. Med. Imaging Graph, 2015.

        [14] Codella, N.C., et al. Skin lesion analysis toward melanoma detection: a challenge at the 2017 international symposium on biomedical imaging (ISBI), hosted by the international skin imaging collaboration (isic). ISBI, 2018.

        [15] Tschandl, P., Rosendahl, C., Kittler, H. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific data, 2018.

        [16] Orlando, J.I., et al. Refuge challenge: a unified framework for evaluating automated methods for glaucoma assessment from fundus photographs. Med. Image Anal, 2020.

        [17] Xinyi Yu, Guanbin Li, Wei Lou, Siqi Liu, Xiang Wan, Yan Chen, and Haofeng Li. Diffusion-Based Data Augmentation for Nuclei Image Segmentation. MICCAI, 2023.

        [18] Kumar, N., et al. A multi-organ nucleus segmentation challenge. IEEE Trans. Med. Imaging, 2019.

        [19] Kumar, N., Verma, R., Sharma, S., Bhargava, S., Vahadane, A., Sethi, A. A dataset and a technique for generalized nuclear segmentation for computational pathology. IEEE Trans. Med. Imaging, 2017.

        [20] Héctor Carrión and Narges Norouzi. FEDD - Fair, Efficient, and Diverse Diffusion-Based Lesion Segmentation and Malignancy Classification. MICCAI, 2023.

        [21] Daneshjou, R., et al. Disparities in dermatology AI performance on a diverse, curated clinical image set. Sci. Adv, 2022.

        [22] Mengxue Sun, Wenhui Huang , and Yuanjie Zheng. Instance-Aware Diffusion Model for Gland Segmentation in Colon Histology Images. MICCAI, 2023.

        [23] Jianfeng Zhao and Shuo Li. Learning Reliability of Multi-modality Medical Images for Tumor Segmentation via Evidence-Identified Denoising Diffusion Probabilistic Models. MICCAI, 2023.

        [24] Baid, U., et al. The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314, 2021.

        [25] Bakas, S., et al. Advancing the cancer genome atlas glioma MRI collections with expert segmentation labels and radiomic features. Sci. Data, 2017.

        [26] Menze, B.H., et al. The multimodal brain tumor image segmentation benchmark (brats). IEEE Trans. Med. Imaging, 2014.

        [27] Jiacheng Wang, Jing Yang, Qichao Zhou, Liansheng Wang. Medical Boundary Diffusion Model for Skin Lesion Segmentation. MICCAI, 2023.

        [28] Gutman, D., et al. Skin lesion analysis toward melanoma detection: A challenge at the international symposium on biomedical imaging (ISBI) 2016, hosted by the international skin imaging collaboration (ISIC). arXiv preprint arXiv:1605.01397, 2016.

        [29] Mendonça, T., Ferreira, P.M., Marques, J.S., Marcal, A.R., Rozeira, J. PH 2-A dermoscopic image database for research and benchmarking. In: 2013 35th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2013.

        [30] Junde Wu, Rao Fu, Huihui Fang, Yu Zhang, Yehui Yang, Haoyi Xiong, Huiying Liu, and Yanwu Xu. MedSegDiff: Medical Image Segmentation with Diffusion Probabilistic Model. MIDL, 2023.

        [31] Fang, H., Li, F., Fu, H., Sun, X., Cao, X., Son, J., Yu, S., Zhang, M., Yuan, C., Bian, C., et al. Refuge2 challenge: Treasure for multi-domain learning in glaucoma assessment. arXiv preprint arXiv:2202.08994, 2022.

        [32] Pedraza, L., Vargas, C., Narváez, F., Durán, O., Muñoz, E., Romero, E. An open access thyroid ultrasound image database. In: 10th International symposium on medical information processing and analysis, 2015.

        [33] Junde Wu, Rao Fu, Huihui Fang, Yu Zhang, and Yanwu Xu. MedSegDiff-V2: Diffusion based Medical Image Segmentation with Transformer. arXiv preprint arXiv:2301.11798, 2023.

        [34] Ji, Y., Bai, H., Yang, J., Ge, C., Zhu, Y., Zhang, R., Li, Z., Zhang, L., Ma, W., Wan, X., et al. Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. arXiv preprint arXiv:2206.08023, 2022.

        [35] Boah Kim, Yujin Oh, Jong Chul Ye. Diffusion Adversarial Representation Learning for Self-supervised Vessel Segmentation. ICLR, 2023.