{"paper":{"title":"Optimization in Sparse 2D to Dense 3D Weakly Supervised Learning: Application to Multi-Label Segmentation of Large ex vivo MRI Data","license":"http://creativecommons.org/licenses/by/4.0/","headline":"2D and 3D segmentation models require distinct regularization when trained from sparse 2D MRI annotations.","cross_cats":["cs.CV","cs.LG"],"primary_cat":"eess.IV","authors_text":"Brandon Bujak, Charidimos Tsagkas, Daniel Reich, Govind Nair, Irene Cortese, Julien Cohen-Adad, Kuan Yi Wang, Paul Hoareau, Roy Sun","submitted_at":"2026-05-12T21:06:53Z","abstract_excerpt":"INTRODUCTION | Fully supervised 3D segmentation of high-resolution ex vivo MRI is limited by the prohibitive cost of volumetric annotation, forcing reliance on sparse 2D slices. Weakly supervised Sparse-to-Dense frameworks bridge this gap, but guidelines remain ambiguous regarding human-centric visual enhancements and transferring optimization strategies across dimensions. We analyze divergent regularization needs for multi-class segmentation of high-resolution ex vivo spinal cord MRI.\n  METHODS | We used 9.4T MRI of multiple sclerosis spinal cords (>104,000 slices) with sparse annotations (42"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"The 2D Teacher required strong spatial augmentation and soft-labeling to overcome data scarcity, improving White Matter Lesion Dice scores by >11 points. However, propagating these techniques to the 3D Student degraded its performance. Furthermore, human-centric preprocessing (e.g., CLAHE) disrupted global statistical cues, dropping Gray Matter Lesion Dice scores by ~25 points.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the pseudo-labels generated by the 2D teacher model are accurate enough to serve as reliable training targets for the 3D student without introducing systematic errors that explain the observed performance differences.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Sparse-to-dense 3D segmentation from 2D slices shows divergent regularization needs: 2D benefits from strong augmentation and soft labels while 3D does not, and human-centric preprocessing harms performance.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"2D and 3D segmentation models require distinct regularization when trained from sparse 2D MRI annotations.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"b7c8ec647bd399121c38773d8f8698e2f013759fac60c2926619c5c1e96f99aa"},"source":{"id":"2605.12753","kind":"arxiv","version":1},"verdict":{"id":"46d2ddb0-2813-4d04-936d-118ce718e9ae","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:32:59.085289Z","strongest_claim":"The 2D Teacher required strong spatial augmentation and soft-labeling to overcome data scarcity, improving White Matter Lesion Dice scores by >11 points. However, propagating these techniques to the 3D Student degraded its performance. Furthermore, human-centric preprocessing (e.g., CLAHE) disrupted global statistical cues, dropping Gray Matter Lesion Dice scores by ~25 points.","one_line_summary":"Sparse-to-dense 3D segmentation from 2D slices shows divergent regularization needs: 2D benefits from strong augmentation and soft labels while 3D does not, and human-centric preprocessing harms performance.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the pseudo-labels generated by the 2D teacher model are accurate enough to serve as reliable training targets for the 3D student without introducing systematic errors that explain the observed performance differences.","pith_extraction_headline":"2D and 3D segmentation models require distinct regularization when trained from sparse 2D MRI annotations."},"references":{"count":24,"sample":[{"doi":"","year":null,"title":"Image Augmentation Techniques for Mammogram Analysis","work_id":"5c1354f1-5c31-492d-acf6-bd7995125de5","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Yoshimi, Yuki and Mine, Yuichi and Ito, Shota and Takeda, Saori and Okazaki, Shota and Nakamoto, Takashi and Nagasaki, Toshikazu and Kakimoto, Naoya and Murayama, Takeshi and Tanimoto, Kotaro. Image p","work_id":"6d48bcf2-7a28-411f-921e-3f2a464ac7a5","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness","work_id":"2d9e320f-8442-4895-b3f6-f360bd2b2289","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Contrastive learning of global and local features for medical image segmentation with limited annotations","work_id":"ec1c032a-b10d-499e-a968-dc5e6df2377b","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Grey matter pathology in multiple sclerosis","work_id":"acb1ff15-956d-4775-9e99-85d11e0ba765","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":24,"snapshot_sha256":"677f87caa42b3b881d0824aa4ef960e47c99a91462714e929d889920ca7185f5","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}