本論文提出 OrthoGW-Fuse（正交灰白質專家融合），灰質專家以 AAL3 區域建構圖，節點特徵取自凍結的自監督三維編碼器同時結合病灶描述先驗，經3視圖的圖注意力網路處理；白質專家則將病灶與 JHU 圖譜的重疊量整理為 eloquent 與 quiet 兩組纖維束分數。三者透過採拉普拉斯近似的貝氏邏輯斯迴歸於決策層融合，使權重維持低維且可直接檢視。於 197 名病患的單中心回溯性世代上，OrthoGW-Fuse 僅以 0.32M 可訓練參數達到AUROC 0.786，並在 AUROC、AUPRC、F1、準確率與敏感度取得所比較基準中最高的值。消融分析顯示，具病灶感知的灰質圖在保有判別力的同時，降低了與白質分支的預測層冗餘（ρ = 0.349，對比無圖結構分支的 0.592），並對應更大的融合增益；橫跨十三個方法下證明我們的方法擁有較低冗餘與較大增益，且校正獨立效能後於 AUROC 仍具顯著性。結果指出，在有限資料下，將灰白質分開建模是善用特徵上非重複的互補性，讓影像分支在雙專家系統中能取得互相補充的重要資訊。

Acute ischemic stroke severity, measured by the NIHSS, reflects the anatomical distribution of tissue injury on diffusion-weighted imaging (DWI). In a small single-center cohort, high-capacity whole-brain models tend to compress gray-matter injury, white-matter involvement, and lesion characteristics into a single score that is hard to interpret and may leave little additional information for a tract-based predictor added afterward. This thesis proposes OrthoGW-Fuse (Orthogonal Gray–White Expert Fusion), which models gray-matter lesion topology and white-matter tract involvement separately and combines them only at the decision level. The gray-matter expert represents each patient as a graph over AAL3 regions, with node features from a frozen self-supervised 3D encoder plus explicit lesion descriptors, processed by a multi-view graph attention network. The white-matter expert summarizes lesion overlap with the JHU atlas into eloquent and quiet tract scores. A Bayesian logistic regression combiner with a Laplace approximation fuses the three scores while keeping the weights low-dimensional and directly inspectable. On a retrospective single-center cohort of 197 patients, OrthoGW-Fuse reaches an AUROC of 0.786 with only 0.32M trainable parameters and obtains the highest point estimates for AUROC, AUPRC, F1, accuracy, and sensitivity among the compared systems. A matched ablation shows that the lesion-aware gray-matter graph retains standalone discrimination while reducing prediction-level redundancy with the tract branch (ρ = 0.349 versus 0.592 for a graph-free branch), yielding a larger fusion gain. Across thirteen methods, lower redundancy is associated with greater gain, and the relationship stays significant for AUROC after adjusting for standalone performance. These results indicate that an image branch's value in a multi-expert system depends on both its standalone signal and the non-duplicated information it adds beyond the other experts, and that separating gray- and white-matter evidence is a practical way to preserve and exploit this complementarity under limited data.

[1] Elizabeth Bates, Stephen M. Wilson, Ayse Pinar Saygin, Frederic Dick, Martin I. Sereno, Robert T. Knight, and Nina F. Dronkers. Voxel-based lesion-symptom mapping. Nature Neuroscience, 6(5):448–450, 2003.
[2] Anna K. Bonkhoff, Alexander L. Cohen, William Drew, Michael A. Ferguson, Aaliya Hussain, Christopher Lin, Frederic L. W. V. J. Schaper, Anthony Bourached, Anne Katrin Giese, Lara C. Oliveira, Robert W. Regenhardt, Markus D. Schirmer, Christina Jern, Arne G. Lindgren, Jane Maguire, Ona Wu, Sahar Zafar, John Y. Rhee, Eyal Y. Kimchi, Maurizio Corbetta, Natalia S. Rost, and Michael D. Fox. Prediction of stroke severity: Systematic evaluation of lesion representations. Annals of Clinical and Translational Neurology, 11(12):3081–3094, 2024.
[3] Anna K. Bonkhoff and Christian Grefkes. Precision medicine in stroke: Towards personalized outcome predictions using artificial intelligence. Brain, 145(2):457–475, 2022.
[4] Anthony Bourached, Anna K. Bonkhoff, Markus D. Schirmer, Robert W. Regenhardt, Martin Bretzner, Sungmin Hong, Adrian V. Dalca, Anne Katrin Giese, Stefan Winzeck, Christina Jern, Arne G. Lindgren, Jane Maguire, Ona Wu, John Rhee, Eyal Y. Kimchi, and Natalia S. Rost. Scaling behaviours of deep learning and linear algorithms for the prediction of stroke severity. Brain Communications, 6(1):fcae007, 2024.
[5] Thomas Brott, Harold P Adams Jr, Charles P Olinger, John R Marler, William G Barsan, Jose Biller, Judith Spilker, Renee Holleran, Robert Eberle, and Vicki Hertzberg. Measurements of acute cerebral infarction: a clinical examination scale. Stroke, 20(7):864–870, 1989.
[6] Sihong Chen, Kai Ma, and Yefeng Zheng. Med3d: Transfer learning for 3d medical image analysis. arXiv preprint arXiv:1904.00625, 2019.
[7] Zijian Chen, Maria Varkanitsa, Prakash Ishwar, Janusz Konrad, Margrit Betke, Swathi Kiran, and Archana Venkataraman. A lesion-aware edge-based graph neural network for predicting language ability in patients with post-stroke aphasia. In International Workshop on Machine Learning in Clinical Neuroimaging, pages 91–101. Springer, 2024.
[8] William A Copen, Albert J Yoo, Natalia S Rost, Lívia T Morais, Pamela W Schaefer, R Gilberto González, and Ona Wu. In patients with suspected acute stroke, ct perfusionbased cerebral blood flow maps cannot substitute for dwi in measuring the ischemiccore. PloS one, 12(11):e0188891, 2017.
[9] Erik Daxberger, Agustinus Kristiadi, Alexander Immer, Runa Eschenhagen, Matthias Bauer, and Philipp Hennig. Laplace redux – effortless bayesian deep learning. In Advances in Neural Information Processing Systems, volume 34, pages 20089–20103, 2021.
[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021.
[11] Valery L. Feigin, Michael Brainin, Bo Norrving, Sheila Martins, Ralph L. Sacco, Werner Hacke, Marc Fisher, Jeyaraj Pandian, and M. Patrice Lindsay. World stroke organization (wso): Global stroke fact sheet 2022. International Journal of Stroke, 17(1):18–29, 2022.
[12] Wuwei Feng, Jasmine Wang, Pratik Y. Chhatbar, Christopher Doughty, Douglas Landsittel, Vasileios-Arsenios Lioutas, Steven A. Kautz, and Gottfried Schlaug. Corticospinal tract lesion load: An imaging biomarker for stroke motor outcomes. Annals of Neurology, 78(6):860–870, 2015.
[13] Larry B. Goldstein, Christina Bertels, and James N. Davis. Interrater reliability of the nih stroke scale. Archives of Neurology, 46(6):660–662, 1989.
[14] R Gilberto Gonzalez. Imaging-guided acute ischemic stroke therapy: from "time is brain" to "physiology is brain". American Journal of Neuroradiology, 27(4):728–735, 2006.
[15] Joseph C. Griffis, Nicholas V. Metcalf, Maurizio Corbetta, and Gordon L. Shulman. Structural disconnections explain brain network dysfunction after stroke. Cell Reports, 28(10):2527–2540.e9, 2019.
[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000–16009, 2022.
[17] Kegang Hua, Jiangyang Zhang, Setsu Wakana, Hangyi Jiang, Xin Li, Daniel S. Reich, Peter A. Calabresi, James J. Pekar, Peter C. M. van Zijl, and Susumu Mori. Tract probability maps in stereotaxic spaces: Analyses of white matter anatomy and tract-specific quantification. NeuroImage, 39(1):336–347, 2008.
[18] Ariel Iporre-Rivas, Dorothee Saur, Karl Rohr, Gerik Scheuermann, and Christina Gillmann. Stroke-gfcn: Ischemic stroke lesion prediction with a fully convolutional graph network. Journal of Medical Imaging, 10(4):044502, 2023.
[19] Hongju Jo, Changi Kim, Dowan Gwon, Jaeho Lee, Joonwon Lee, Kang Min Park, and Seongho Park. Combining clinical and imaging data for predicting functional outcomes after acute ischemic stroke: An automated machine learning approach. Scientific Reports, 13(1):16926, 2023.
[20] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations, 2017.
[21] Anders Krogh and Jesper Vedelsby. Neural network ensembles, cross validation, and active learning. In Advances in Neural Information Processing Systems, volume 7, pages 231–238. 1995.
[22] Bo Mi Kwon, Jin-Youn Lee, Nayeon Ko, Bo-Ram Kim, Won-Jin Moon, Dong-Hee Choi, and Jongmin Lee. Correlation of hemispatial neglect with white matter tract integrity: A dti study. Brain & NeuroRehabilitation, 15(1):e6, 2022.
[23] Xiaoxiao Li, Yuan Zhou, Nicha Dvornek, Muhan Zhang, Siyuan Gao, Juntang Zhuang, Dustin Scheinost, Lawrence H. Staib, Pamela Ventola, and James S. Duncan. Braingnn: Interpretable brain graph neural network for fmri analysis. Medical Image Analysis, 74:102233, 2021.
[24] Paul Pu Liang, Yun Cheng, Xiang Fan, Chun Kai Ling, Suzanne Nie, Richard J. Chen, Zihao Deng, Nicholas Allen, Randy Auerbach, Faisal Mahmood, Ruslan Salakhutdinov, and Louis-Philippe Morency. Quantifying and modeling multimodal interactions: An information decomposition framework. In Advances in Neural Information Processing Systems, volume 36, pages 27351–27393, 2023.
[25] Mingtian Liu, Nima Hatami, Laura Mechtouff, Tae-Hee Cho, Carole Lartizien, and Carole Frindel. Lesion-centered vision transformer for stroke outcome prediction from image and clinical data. In Medical Image Computing and Computer Assisted Intervention, volume 15974, pages 291–301. Springer Nature Switzerland, 2025.
[26] Yongkai Liu, Yannan Yu, Jiahong Ouyang, Bin Jiang, Guang Yang, Sophie Ostmeier, Max Wintermark, Patrik Michel, David S. Liebeskind, Maarten G. Lansberg, Gregory W. Albers, and Greg Zaharchuk. Functional outcome prediction in acute ischemic stroke using a fused imaging and clinical deep learning model. Stroke, 54(9):2316–2327, 2023.
[27] Xuexiong Luo, Jia Wu, Jian Yang, Shan Xue, Amin Beheshti, Quan Z. Sheng, David McAlpine, Paul Sowman, Alexis Giral, and Philip S. Yu. Graph neural networks for brain graph learning: A survey. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 8170–8178, 2024.
[28] David J. C. MacKay. The evidence framework applied to classification networks. Neural Computation, 4(5):720–736, 1992.
[29] Hamed Mohammadi and Waldemar Karwowski. Graph neural networks in brain connectivity studies: Methods, challenges, and future directions. Brain Sciences, 15(1):17, 2025.
[30] Seunghun Oh, Hyunsu Jeong, Yoonjae Cho, Kiduk Kim, Chulhong Kim, Dong Young Jeong, and Bum Joon Kim. Mixture-of-skip-connection deep learning model to classify stroke severity from diffusion weighted imaging based on nihss. Journal of Imaging Informatics in Medicine, 2026. Published online 6 March 2026.
[31] Josep Puig, Salvador Pedraza, Gerard Blasco, Josep Daunis-i Estadella, Ferran Prados, Sebastià Remollo, Albert Prats-Galino, Guadalupe Soria, Imma Boada, Mar Castellanos, and Joaquín Serena. Acute damage to the posterior limb of the internal capsule on diffusion tensor tractography as an early imaging predictor of motor outcome after stroke. American Journal of Neuroradiology, 32(5):857–863, 2011.
[32] Deepthi Rajashekar, Matthias Wilms,M Ethan MacDonald, Serena Schimert, Michael D Hill, Andrew Demchuk, Mayank Goyal, Sean P Dukelow, and Nils Daniel Forkert. Lesion-symptom mapping with nihss sub-scores in ischemic stroke patients. Stroke and vascular neurology, 7(2), 2022.
[33] Alessandro Salvalaggio, Michele De Filippo De Grazia, Marco Zorzi, Michel Thiebaut de Schotten, and Maurizio Corbetta. Post-stroke deficit prediction from lesion and indirect structural and functional disconnection. Brain, 143(7):2173–2188, 2020.
[34] Pamela W Schaefer, Wiliam A Copen, Michael H Lev, and R Gilberto Gonzalez. Diffusion-weighted imaging in acute stroke. Magnetic Resonance Imaging Clinics, 14(2):141–168, 2006.
[35] Christoph Sperber, Laura Gallucci, Vanessa Kasties, Marcel Arnold, and Roza M Umarova. Connectomic stroke lesion measures provide no benefit over basic spatial lesion features in the prognosis of global stroke outcome measures. Brain Communications, 7(4):fcaf268, 2025.
[36] Christoph Sperber, Laura Gallucci, Daniel Mirman, Marcel Arnold, and Roza M. Umarova. Stroke lesion size – still a useful biomarker for stroke severity and outcome in times of high-dimensional models. NeuroImage: Clinical, 40:103511, 2023.
[37] Yucheng Tang, Dong Yang, Wenqi Li, Holger R. Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swin transformers for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20730–20740, 2022.
[38] Michel Thiebaut de Schotten, Chris Foulon, and Parashkev Nachev. Brain disconnections link structural connectivity with function and behaviour. Nature Communications, 11(1):5094, 2020.
[39] Chia-Ling Tsai, Hui-Yun Su, Shen-Feng Sung, Wei-Yang Lin, Ying-Ying Su, Tzu-Hsien Yang, and Man-Lin Mai. Fusion of diffusion-weighted mri and clinical data for predicting functional outcome after acute ischemic stroke with deep contrastive learning. arXiv preprint arXiv:2402.10894, 2024.
[40] Pahati Tuxunjiang, Chencui Huang, Zhen Zhou, Wenyi Zhao, Bingyan Han, Weixiong Tan, Jingru Wang, HuangHanjiaerbieke Kukun, Wei Zhao, Rui Xu, Ainikaerjiang Aihemaiti, Yimuran Subi, Jingyang Zou, Chao Xie, Yifan Chang, and Yunling Wang. Prediction of nihss scores and acute ischemic stroke severity using a cross-attention vision transformer model with multimodal mri. Academic Radiology, 32(9):5453–5467, 2025.
[41] Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017.
[42] Emilia Vitti, Ganghyun Kim, Melissa D. Stockbridge, Argye E. Hillis, and Andreia V. Faria. Left hemisphere bias of nih stroke scale is most severe for middle cerebral artery strokes. Frontiers in Neurology, 13:912782, 2022.
[43] Tassilo Wald, Constantin Ulrich, Stanislav Lukyanenko, Andrei Goncharov, Alberto Paderno, Maximilian Miller, Leander Maerkisch, Paul Jaeger, and Klaus Maier-Hein. Revisiting mae pre-training for 3d medical image segmentation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 5186–5196, 2025.
[44] Danny Wood, Tingting Mu, Andrew M. Webb, Henry W. J. Reeve, Mikel Luján, and Gavin Brown. A unified theory of diversity in ensemble learning. Journal of Machine Learning Research, 24(359):1–49, 2023.
[45] Ona Wu, Lisa Cloonan, Steven J. T. Mocking, Mark J. R. J. Bouts, William A. Copen, Pedro Cougo-Pinto, Kaitlin Fitzpatrick, Allison Kanakis, Pamela W. Schaefer, Jonathan Rosand, Karen L. Furie, and Natalia S. Rost. Role of acute lesion topography in initial ischemic stroke severity and long-term functional outcomes. Stroke, 46(9):2438–2444, 2015.
[46] Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? In International Conference on Learning Representations, 2019.
[47] Ying Zeng, Chen Long, Wei Zhao, and Jun Liu. Predicting the severity of neurological impairment caused by ischemic stroke using deep learning based on diffusion-weighted images. Journal of Clinical Medicine, 11(14):4008, 2022.
[48] Yongsheng Zhang, Daniel Y. Kimberg, H. Branch Coslett, Myrna F. Schwartz, and Ze Wang. Multivariate lesion-symptom mapping using support vector regression. Human Brain Mapping, 35(12):5861–5876, 2014.
[49] Lin L. Zhu, Robert Lindenberg, Michael P. Alexander, and Gottfried Schlaug. Lesion load of the corticospinal tract predicts motor impairment in chronic stroke. Stroke, 41(5):910–915, 2010.