{"paper_id":"a74debdf-9e5e-4328-80d9-a5dacf97102f","body_text":"DISTILLING MISSING MODALITY KNOWLEDGE FROM ULTRASOUND FOR\nENDOMETRIOSIS DIAGNOSIS WITH MAGNETIC RESONANCE IMAGES\nYuan Zhang1, Hu Wang1, David Butler 1, Minh-Son To2, Jodie Avery3, M Louise Hull 3, Gustavo Carneiro4\n1 Australian Institute for Machine Learning, University of Adelaide, Australia\n2 Flinders Health and Medical Research Institute, Flinders University, Australia\n3 Robinson Research Institute, University of Adelaide, Australia\n4 Centre for Vision, Speech and Signal Processing, University of Surrey, UK\nABSTRACT\nEndometriosis is a common chronic gynecological disorder\nthat has many characteristics, including the pouch of Douglas\n(POD) obliteration, which can be diagnosed using Transvagi-\nnal gynecological ultrasound (TVUS) scans and magnetic res-\nonance imaging (MRI). TVUS and MRI are complementary\nnon-invasive endometriosis diagnosis imaging techniques,\nbut patients are usually not scanned using both modalities\nand, it is generally more challenging to detect POD obliter-\nation from MRI than TVUS. To mitigate this classification\nimbalance, we propose in this paper a knowledge distillation\ntraining algorithm to improve the POD obliteration detection\nfrom MRI by leveraging the detection results from unpaired\nTVUS data. More specifically, our algorithm pre-trains a\nteacher model to detect POD obliteration from TVUS data,\nand it also pre-trains a student model with 3D masked auto-\nencoder using a large amount of unlabelled pelvic 3D MRI\nvolumes. Next, we distill the knowledge from the teacher\nTVUS POD obliteration detector to train the student MRI\nmodel by minimizing a regression loss that approximates the\noutput of the student to the teacher using unpaired TVUS and\nMRI data. Experimental results on our endometriosis dataset\ncontaining TVUS and MRI data demonstrate the effective-\nness of our method to improve the POD detection accuracy\nfrom MRI.\nIndex Terms— Endometriosis, Knowledge Distillation,\nMasked Auto-Encoder, Pouch of Douglas Obliteration\n1. INTRODUCTION\nEndometriosis is a gynecological disorder associated with the\ngrowth of endometrial glands and stroma outside the uter-\nine cavity [1, 2]. The clinical manifestations of endometrio-\nsis include infertility and endometriosis-related pain [3]. As\na common chronic gynecological disease, it affects approxi-\nmately 1.5 million women worldwide [4]. There is currently\nno known way to prevent or cure endometriosis, but early di-\nagnosis, intervention and management may slow or stop the\nnatural disease progression. However, the diagnosis of en-\ndometriosis can take about 7 years on average after the ap-\npearance of initial symptoms [5]. Laparoscopy used to be the\ndiagnostic gold standard [6], but with the improvement in the\nquality and availability of imaging modalities for endometrio-\nsis diagnosis, there has been evidence suggesting that accu-\nrate endometriosis diagnosis can be achieved with the analy-\nsis of TVUS sequences and MRI volumes [7, 8].\nMany of the symptomatic endometriosis cases can be as-\nsociated with the pouch of Douglas (POD) obliteration, which\ncan be diagnosed from complementary TVUS and MRI data,\nas shown in Fig. 1. However, in clinical practice, it is difficult\nto access clinicians who can diagnose endometriosis with one\nof these modalities, not to mention those who are proficient\nin both modalities. For TVUS, POD obliteration can be ac-\ncurately detected manually [9] and automatically [10] via the\nultrasound ‘sliding sign’ [11], which is classified as positive\n(i.e. normal) or negative (i.e. abnormal), where a negative\nsliding sign is considered when the anterior rectum or bowel\nglides cannot freely slide over the posterior upper uterine\nfundus. For MRI, POD can be classified as obliterated or\nnormal, where the POD obliteration can be characterized by\nendometrial plaques and dense adhesions between uterosacral\nligaments, uterine serosa, ovaries, rectum and vagina on T2-\nweighted and T1-weighted images [12]. However, differently\nfrom TVUS, the manual POD obliteration detection from\nMRI can only reach 61.4-71.9% accuracy [13]. There has\nbeen some effort to propose methods that can automatically\ndiagnose deep pelvic endometriosis classification from MRI1,\nbut we are not aware of methods that can detect POD obliter-\nation from MRI.\nLeveraging the TVUS POD obliteration detection to im-\nprove the automated detection accuracy from MRI using an\nunpaired training set containing scans from both modalities is\nthe main goal of this paper. We achieve this goal by proposing\na new knowledge distillation algorithm based on two stages:\n1) pre-training a teacher model to detect POD obliteration\nfrom TVUS data, and pre-training a student model with 3D\nmasked auto-encoder (MAE) using a large amount of unla-\nbelled pelvic 3D MRI volumes; and 2) knowledge distilla-\ntion from the teacher TVUS detector to train the student MRI\nmodel by minimizing a regression loss that approximates the\noutput of the student to the teacher using unpaired TVUS and\n1Deep infiltrating endometriosis (DIE) can lead to a partial or complete\nobliteration of the pouch of Douglas (POD) [14].\narXiv:2307.02000v1  [eess.IV]  5 Jul 2023\n\nFig. 1 . Examples of POD obliteration on MRI and sliding\nsign on TVUS. (a) and (b) represent negative and positive\nPOD obliteration sign on sagittal plane MRI, respectively. (c)\nand (d) represent positive and negative sliding sign on TVUS,\nrespectively.\nMRI data. The testing is realised using only MRI data. The\nmain innovations of this work are:\n• To the best of our knowledge, this is the first POD oblitera-\ntion detection method that distills knowledge from TVUS\nto MRI using unpaired data, with the objective of improv-\ning the accuracy of diagnosing endometriosis from MRI;\n• It is also the first machine learning method that can auto-\nmatically detect POD obliteration from MRI data with the\ngoal of diagnosing endometriosis.\nExperimental results on a private endometriosis training set\ncontaining unpaired TVUS and MRI data show the effective-\nness of our method to increase the POD detection accuracy\nfrom testing MRI volumes.\n2. RELA TED WORK\nThe automated detection of endometriosis from medical\nimaging has received some attention lately. Using ultrasound\nimages, Guerriero et al. [15] compared the ability of six ma-\nchine learning algorithms and neural networks for the diag-\nnosis of endometriosis in the rectosigmoid, where the neural\nnetwork achieved the highest classification accuracy of 0.73.\nMaicas et al. [10] constructed a deep learning model based\non a temporal residual network to classify POD obliteration\nfrom TVUS videos, achieving an AUC of 96.5% and an accu-\nracy of 88.8%. The methods above use TVUS data alone, but\nrecently, Yang et al. [16] built a bi-model method with one\nmodel for MRI and another for TVUS, but they do not ex-\nplore the relationship between MRI and TVUS, like we pro-\npose in this paper. Furthermore, as mentioned above, it is\nunlikely that patients will have access to both modalities in\nclinical practice, which justifies the need for single-modality\napproaches that have highly accurate endometriosis detection.\nKnowledge distillation is a general framework to extract\nthe knowledge learned from a teacher model to a student\nmodel by soft-label supervision. The original purpose of\nknowledge distillation is to compress deep learning mod-\nels, so they can run on resource-constrained devices [17],\nbut in this paper we focus on the transfer of knowledge\nfrom a teacher network trained on source modalities to a\nstudent network that is trained on different target modali-\nties [18]. In medical image analysis, data from different\nmodalities can provide rich and complementary information\nabout diseases, so multimodal knowledge distillation is suit-\nable for scenarios where data or labels for some modalities\nare missing during training or testing. Inspired by knowl-\nedge distillation, Dou et al. [19] tackled the task of unpaired\nmulti-modal image segmentation. Guan et al. [20] leverage\nadditional supervision distilled from clinical data to improve\nMRI-based Alzheimer’s disease prediction models. Cheng et\nal. [21] utilise both a pixel-level and an image-level distilla-\ntion scheme to transfer knowledge from a multimodal-MRI\nteacher network to a unimodal segmentation student network.\nHowever, most unpaired multi-modal studies above focus on\nMRI and CT scans, which is arguably easier than focusing on\nMRI and TVUS, which is the case considered in this paper.\n3. METHOD\nAn overview of our proposed TVUS to MRI knowledge distil-\nlation model is shown in Figure 2, which consists of two mod-\nels: a teacher model pre-trained with a TVUS dataset and a\nstudent model pre-trained on an MRI dataset, and then trained\non an unpaired dataset of TVUS and MRI data by distilling\nthe knowledge in the representation learned by the teacher\nmodel to the student model.\nFormally, letDM = {(xi, yi)}N\ni=1 denote the endometrio-\nsis MRI dataset, with N T2 SPC images x ∈ X ⊂ RH×W ×D\nand corresponding POD obliteration one-hot label y ∈\n{0, 1}2, where H, W and D are height, width and depth\nof the MRI, respectively. For TVUS dataset, let Ds\nU =\n{(xs\ni , ys\ni )}M\ni=1 be the video clips dataset, where xs ∈ X s ⊂\nRH×W ×T (H, W and T are height, width and number of\nframes), and ys ∈ {0, 1}2 denotes the POD obliteration one-\nhot label. For the self-supervised pre-training of the MRI\nPOD obliteration detector, we have DP = {xp\ni }K\ni=1, which\ncontains K unlabeled MRI volumes, where the number of\nunlabeled images is much larger than the labeled images (i.e.,\nK >> N and K >> M ). The teacher model fθU : X s → ∆\n(with ∆ ⊂ [0, 1]2 being the probability simplex) is trained\nwith dataset Ds\nU , the student model fθM : X → ∆ is pre-\ntrained with dataset DP and fine-tuned using DM . The final\nknowledge distillation model is initialised by the pre-trained\nstudent model fθM (.), which is then trained from DM and\nDs\nU . The testing to classify POD obliteration uses the trained\nfθM (.) on the MRI testing images from DM .\n3.1. Pre-training\nFor pre-training the MRI encoder, we explore the self-\nsupervised masked auto-encoder (MAE) method [22] us-\ning the large dataset of un-annotated T2 MRI images DP .\n\nFig. 2 . Proposed POD obliteration detector trained by dis-\ntilling knowledge to MRI from unpaired TVUS. (a) MRI pre-\ntraining with 3D masked auto-encoder, (b) TVUS pre-training\nwith ResNet(2+1)D, (c) MRI Knowledge Distillation from\nthe frozen teacher model pretrained on TVUS.\n3D-MAE has an asymmetric encoder-decoder architecture\nbased on 3D Vision Transformer (3D ViT) [23], as shown in\nFigure 2. During pre-training, each volume is cropped into\n8 × 8 × 8 patches, then 50% of the patches are randomly\nmasked. The encoder, denoted by fθE : X → E , takes visible\npatches embedded by a linear projection (of dimension 768)\nwith additional positional embeddings, and processes via the\n3D ViT with 12 Transformer blocks whose output is then fed\ninto the decoder, denoted by fθD : D → X , together with the\nmasked volume tokens, to reconstruct the original volume at\nthe voxel level. After pre-training, the decoder is discarded\nand the pre-trained encoder is applied to extract MRI fea-\ntures for the downstream classification task. During MRI\npre-training, we minimize the mean squared error (MSE) on\nthe masked patches between the reconstructed and original\nvolumes in the voxel space, defined as follows:\nℓM AE(DP ; θE, θD) = 1\nL × K\nKX\ni=1\nLX\nl=1\n∥xi(l) − ˆxi(l)∥2\n2,\n(1)\nwhere K is the size of the unlabelled MRI dataset,L indicates\nthe number of masked patches, xi(l) and ˆxi(l) represent the\nvoxel values of thelth masked patch in the original and recon-\nstructed volumes, respectively, with the reconstructed volume\nobtained from ˆxi = fθD(fθE(xi)).\nAfter this self-supervised pre-training, we take the en-\ncoder fθE(.), add a linear layer to change the size from 768\n(output size of the MRI encoder) to 512 (output size of the\nTVUS encoder), and add a final classification layer with a 2-\ndimensional output activated by softmax to form the student\nnetwork, denoted by fθM (.). We fine-tune all transformer\nblocks in the encoder and fully connected layer to classify\nPOD obliteration by minimizing the cross-entropy (CE) loss:\nℓP T M(DM; θM) = −\nX\n(xi,yi)∈DM\nℓCE(yi, fθM (xi)). (2)\nFor the TVUS pre-training, we adopted the ResNet\n(2+1)D model proposed in [10]. This model contains 18\nmodules of R(2+1) convolutional layers, with each convo-\nlution being followed by batch normalization. During this\nTVUS pre-training, we also minimize the CE loss, as fol-\nlows:\nℓP T U(Ds\nU; θU) = −\nX\n(xs\ni ,ys\ni )∈DM\nℓCE(ys\ni , fθU (xs\ni )). (3)\n3.2. Knowledge Distillation\nIn the knowledge distillation (KD) stage, we consider the pre-\ntrained TVUS model fθU (.) as the teacher model, and the pre-\ntrained MRI model fθM (.) as the student model. Given that\nthe pre-trained TVUS model tends to produce superior classi-\nfication accuracy than the pre-trained MRI model, we aim to\nfine-tune the MRI model to match the predictions produced\nby the TVUS model. The goal is to use this knowledge distil-\nlation procedure to improve the classification accuracy of the\nMRI model, which uses only the MRI volume during testing.\nHowever, recall that we do not have matched TVUS and MRI\ndata for this knowledge distillation procedure, so we match\nthe data based only on their classification labels, i.e., an MRI\nsample with positive POD obliteration is matched with a ran-\ndom TVUS sample with positive POD obliteration, and simi-\nlarly for the negative POD obliteration. Then, the KD training\nminimises the following loss\nℓKD(DM , Ds\nU; θM) = −\nX\n(xi,yi)∈DM\n(xs\nj ,ys\nj )∈Ds\nU\nyi=ys\nj\n∥fθU (xs\nj) − fθM (xi))∥1.\n(4)\nThe loss in (4) is added to the ℓP T M(.) loss from (2) to form\nthe final loss that encourages the model to pull the TVUS and\nthe MRI outputs closer to distill the TVUS classification in-\nformation to the student network, as follows\nℓ(DM , Ds\nU; θM) =αepoch × ℓKD(DM , Ds\nU; θM)+\n(1 − αepoch) × ℓP T M(DM; θM),\n(5)\nwhich is used to estimate the optimal value of θM , where\nαepoch is a parameter to dynamically balance the contribu-\ntions of the two loss terms during training.\n4. EXPERIMENTS\n4.1. Dataset\nOur private dataset contains: an MRI endometriosis dataset,\na TVUS endometriosis dataset, and a female pelvic MRI\ndataset for self-supervised pre-training.\nThe MRI endometriosis dataset contains 88 T2 SPACE\nMRI scans from women aged 18-45, including 19 POD oblit-\neration cases. These examinations were performed in several\nclinical sites in Australia. The scans contain the whole pelvic\narea, but we focus on the region around the uterus that can\n\ndisplay the POD obliteration. We performed 3D contrast lim-\nited adaptive histogram equalization (CLAHE) to improve the\nlocal contrast and enhance the definitions of edges in each re-\ngion of a sequence. For the experiments, we use stratified\nrandom sampling to split the dataset into 5 folds, each with\n17-18 subjects, with 4 folds used for training and 1 for test-\ning, which are rotated to produce a cross-validated result.\nThe TVUS endometriosis dataset has 749 video clips of\nthe ’sliding sign’, including 103 negative (POD obliterated)\ncases. We follow the data preparation and pre-processing pro-\ntocol as well as model parameter settings proposed in [10]\nto pre-train the TVUS model. For the knowledge distillation\nphase, we divide the dataset into negative and positive groups,\nthen use stratified random sampling to split each group into 5\nfolds and use 4 folds in each group as the training set.\nThe female pelvic MRI dataset contains 8,984 volumes.\nIn the context of endometriosis research, we accept all scans\nwhen the patient is aged between 18 to 45, the physical ex-\namination site is pelvis, and the sequence description con-\ntains ’T2’. It is worth noting that most of these volumes were\nscanned in different settings, so they may contain signs of\nother diseases and the scanned area may or may not overlap\nwith the diagnostic area of endometriosis. We re-sampled all\nscans by SimpleITK with an output spacing of 1 × 1 × 1mm,\nand filtered out volumes if the number of slices in any dimen-\nsion is less than 65.\n4.2. Training and Evaluation\nIn the pre-training phase, the 3D MAE network is trained for\n200 epochs on the female pelvic MRI dataset using a batch\nsize of 3 with AdamW optimizer and a learning rate of 1e-\n3. Then we fine-tuned the pre-trained checkpoint saved from\nepoch 195 for 25 epochs on the endometriosis MRI dataset\nwith a “5-fold cross validation” strategy. The ResNet(2+1)D\nnetwork was pre-trained on the Kinetics-400 dataset then fine-\ntuned for 30 epochs on the TVUS endometriosis dataset using\na batch size of 30 with Adam optimizer and a learning rate of\n1e-5. For knowledge distillation, we loaded the weights of the\n10th checkpoint fine-tuned from each fold as the MRI feature\nextractor of the MRI model and the pre-trained TVUS feature\nextractor. The knowledge distillation network is trained for 10\nepochs on the endometriosis MRI dataset using a batch size\nof 7 with AdamW optimizer and a learning rate of 1e-3 with\nthe same “5-fold cross validation” strategy. We set the hyper-\nparameter α = 0.85 based on the cross-validation results. We\nevaluate our method by computing Area Under the Receiver\nOperating Characteristic Curve (ROC AUC) for five folds and\ncalculate the mean and standard deviation of them.\n4.3. Results\nThe classification results with our proposed method is shown\nin Table 1. The ResNet(2+1)D model shows an outstanding\nclassification AUC of 96.9% from TVUS data. The small\namount of training samples limited the generalisation of 3D\nViT to classify POD obliteration from MRI volumes with an\nAUC of 65.0%, but the MAE pre-training (PT) partially mit-\nigates the issue, improving the AUC to 87.2%. With knowl-\nedge distillation (KD), we observe that training a 3D ViT from\nscratch on such a small dataset is still challenging, with an\nAUC of 66.7%. Also, the KD performance of 3D ViT with\nMAE pre-training (PT) reaches AUC=77.2%), which is worse\nthan without KD, with AUC=87.2%, which may be due to the\nexcessive domain shift between the pre-training dataset and\nTVUS dataset. By fine-tuning (FT) the model from MAE\npre-training, the model improves accuracy from AUC=87.2%\n(without KD and FT) to AUC=90.6% (with KD and FT).\nTable 1. POD obliteration classification results.Method Training Testing AUC\nModality Modality mean±stddev\nResNet(2+1)D TVUS TVUS 0.969±0.012\n3D ViT MRI MRI 0.650±0.102\n3D ViT: MAE PT MRI MRI 0.872±0.094\n3D ViT: KD MRI,TVUSMRI 0.667±0.107\n3D ViT: MAE PT + KD MRI,TVUSMRI 0.772±0.087\n3D ViT: MAE PT + KD + FTMRI,TVUSMRI 0.906±0.099\n5. CONCLUSION\nIn this paper, we proposed a two-stage algorithm to distill\nthe knowledge from a TVUS to an MRI classifier, thereby\nimproving the POD obliteration classification accuracy of\nthe MRI classifier. Through the MAE pre-training, knowl-\nedge distillation and fine-tuning, we are able to significantly\nreduce the distance between the two domains and accom-\nplish a promising knowledge distillation from TVUS to MRI.\nThe efficacy and superiority of our proposed approach are\ndemonstrated by experimental results on our endometriosis\ndatasets. In the future, we will introduce a missing modality\ndeep learning approach and expand our proposed method to\nperform weakly-supervised lesion segmentation, thereby im-\nproving the interpretability of the model, so it can be widely\napplied in future clinical trials.\n6. COMPLIANCE WITH ETHICAL STANDARDS\nThis study was performed in line with the principles of the\nDeclaration of Helsinki. Approval was granted by Human\nResearch Ethics Committee (HREC) of University of Ade-\nlaide(Date 01-03-2020/No. H-2020-051) and the Southern\nAdelaide Clinical Human Research Ethics Committee (SAC\nHREC) (Date 01-11-2021/No. 111.20).\n7. ACKNOWLEDGMENTS\nThis work received funding from the Australian Government\nthrough the Medical Research Futures Fund: Primary Health\nCare Research Data Infrastructure Grant 2020 and from En-\ndometriosis Australia.\n\n8. REFERENCES\n[1] Antonio Simone Lagana et al., “Evaluation of m1\nand m2 macrophages in ovarian endometriomas from\nwomen affected by endometriosis at different stages of\nthe disease,” Gynecological Endocrinology, vol. 36, no.\n5, pp. 441–444, 2020.\n[2] KM Moss, J Doust, H Homer, IJ Rowlands, R Hockey,\nand GD Mishra, “Delayed diagnosis of endometriosis\ndisadvantages women in art: a retrospective population\nlinked data study,” Human Reproduction, vol. 36, no.\n12, pp. 3074–3082, 2021.\n[3] Paolo Vercellini et al., “Endometriosis: pathogenesis\nand treatment,” Nature Reviews Endocrinology, vol. 10,\nno. 5, pp. 261–275, 2014.\n[4] Krina T. Zondervan et al., “Endometriosis,” New Eng-\nland Journal of Medicine , vol. 382, no. 13, pp. 1244–\n1256, 2020, PMID: 32212520.\n[5] Australian Institute of Health and Welfare, “En-\ndometriosis in australia: prevalence and hospitalisa-\ntions,” 2019, [Online; accessed 18-02-2022].\n[6] Christian M Becker et al., “Eshre guideline: en-\ndometriosis,” Human reproduction open, vol. 2022, no.\n2, pp. hoac009, 2022.\n[7] Alison Deslandes et al., “Current status of transvaginal\nultrasound accuracy in the diagnosis of deep infiltrating\nendometriosis before surgery: a systematic review of the\nliterature,” Journal of Ultrasound in Medicine , vol. 39,\nno. 8, pp. 1477–1490, 2020.\n[8] T Indrielle-Kelly et al., “Diagnostic accuracy of ul-\ntrasound and mri in the mapping of deep pelvic en-\ndometriosis using the international deep endometriosis\nanalysis (idea) consensus,” BioMed research interna-\ntional, vol. 2020, 2020.\n[9] Le Chi Chiu et al., “Predicting pouch of douglas oblit-\neration using ultrasound and laparoscopic video sets: an\ninterobserver and diagnostic accuracy study,”Journal of\nUltrasound in Medicine, vol. 38, no. 12, pp. 3155–3161,\n2019.\n[10] Gabriel Maicas et al., “Deep learning to diagnose pouch\nof douglas obliteration with ultrasound sliding sign,”\nReproduction and Fertility , vol. 2, no. 4, pp. 236–243,\n2021.\n[11] Stefano Guerriero et al., “Systematic approach to sono-\ngraphic evaluation of the pelvis in women with sus-\npected endometriosis, including terms, definitions and\nmeasurements: a consensus opinion from the interna-\ntional deep endometriosis analysis (idea) group,” Ultra-\nsound in Obstetrics & Gynecology , vol. 48, no. 3, pp.\n318–332, 2016.\n[12] Anitha L Thalluri et al., “Mri findings in deep infiltrat-\ning endometriosis: a pictorial essay,” Journal of Medi-\ncal Imaging and Radiation Oncology , vol. 61, no. 6, pp.\n767–773, 2017.\n[13] Milliam L Kataoka et al., “Posterior cul-de-sac obliter-\nation associated with endometriosis: Mr imaging evalu-\nation,” Radiology, vol. 234, no. 3, pp. 815–823, 2005.\n[14] Kristina Arion et al., “Prediction of pouch of douglas\nobliteration: point-of-care ultrasound versus pelvic ex-\namination,” Journal of Minimally Invasive Gynecology ,\nvol. 26, no. 5, pp. 928–934, 2019.\n[15] Stefano Guerriero et al., “Artificial intelligence (ai) in\nthe detection of rectosigmoid deep endometriosis,” Eu-\nropean Journal of Obstetrics & Gynecology and Repro-\nductive Biology, vol. 261, pp. 29–33, 2021.\n[16] Minmin Yang et al., “Diagnostic efficacy of ultrasound\ncombined with magnetic resonance imaging in diag-\nnosis of deep pelvic endometriosis under deep learn-\ning,” The Journal of Supercomputing , vol. 77, no. 7,\npp. 7598–7619, 2021.\n[17] Geoffrey Hinton et al., “Distilling the knowledge in a\nneural network,” arXiv preprint arXiv:1503.02531, vol.\n2, no. 7, 2015.\n[18] Nikolaos Passalis et al., “Learning deep representa-\ntions with probabilistic knowledge transfer,” in Pro-\nceedings of the European Conference on Computer Vi-\nsion (ECCV), 2018, pp. 268–284.\n[19] Qi Dou et al., “Unpaired multi-modal segmentation via\nknowledge distillation,” IEEE transactions on medical\nimaging, vol. 39, no. 7, pp. 2415–2425, 2020.\n[20] Hao Guan et al., “Mri-based alzheimer’s disease predic-\ntion via distilling the knowledge in multi-modal data,”\nNeuroImage, vol. 244, pp. 118586, 2021.\n[21] Cheng Chen et al., “Learning with privileged multi-\nmodal knowledge for unimodal segmentation,” IEEE\nTransactions on Medical Imaging , vol. 41, no. 3, pp.\n621–632, 2021.\n[22] Kaiming He et al., “Masked autoencoders are scal-\nable vision learners,” arXiv preprint arXiv:2111.06377,\n2021.\n[23] Alexey Dosovitskiy et al., “An image is worth 16x16\nwords: Transformers for image recognition at scale,”\narXiv preprint arXiv:2010.11929, 2020.","source_license":"CC0","license_restricted":false}