{
  "task_slug": "mri-2slice-observation-geometry",
  "question": "在严格等观测预算(2个切片)下,正交双切片(矢状+冠状)相比平行双切片(2个矢状,间距d),能否显著降低实时MRI引导放疗中低秩运动模型驱动的三维形变场重建误差?误差差异由哪条物理机制(穿层位移不可观测)解释?",
  "falsifiable_prediction": "K=3, sigma_sag=0.5mm, alpha=1,非交错默认设定下:正交臂相对最佳间距平行矢状臂的dvf_rmse3d_mm改善<0.5mm且<20%相对值,27个配对样本中方向为正的<=19个;同时平行矢状臂的rmse_lr_mm>=正交臂的1.5倍,但LR方向贡献的平方误差占比<15%。推翻条件:(a)正交改善>=0.5mm且>=20%且bootstrap95%CI下界>0;(b)平行矢状以同等幅度反超;(c)平行矢状rmse_lr_mm不到正交的1.5倍。",
  "prediction_outcome": "refuted",
  "negative_result": true,
  "dataset": "ETH Zurich CVL 4DMRI 公开集 (von Siebenthal et al., MICCAI 2005), motionfields/motionfield_subj{A,B}_cyc{n}_ds5.txt, 5mm各向同性网格点轨迹; subject A: 8219点 x 4呼吸周期(16/16/19/?帧), subject B: 10125点 x 5呼吸周期。位移相对各受试者cyc3第0帧参考位置定义。留一周期交叉验证(9折 = A的4折+B的5折)x3种子=27配对样本。",
  "env": {
    "python": "3.12",
    "torch": "2.13.0+cpu",
    "extra_packages": []
  },
  "seeds": [
    0,
    1,
    2
  ],
  "arms": [
    {
      "name": "baseline-single-sagittal",
      "is_baseline": true,
      "what": "1 个矢状切片,过 ROI 中心,层厚5mm(等效层内5mm网格层),LR方向无观测,半预算参照。",
      "metrics": {
        "dvf_rmse3d_mm": {
          "per_seed": [
            0.7608870824421163,
            0.7603149348587664,
            0.7633959730031891
          ],
          "mean": 0.7615326634346906,
          "std": 0.001338103166258539
        },
        "dvf_p95_mm": {
          "per_seed": [
            1.5572315235835799,
            1.5540512346150388,
            1.5630976760641702
          ],
          "mean": 1.5581268114209295,
          "std": 0.0037470593765351755
        },
        "rmse_ap_mm": {
          "per_seed": [
            0.34613450982139515,
            0.346113218680625,
            0.3469521625867928
          ],
          "mean": 0.34639996369627096,
          "std": 0.00039056031475216094
        },
        "rmse_si_mm": {
          "per_seed": [
            0.31914009890056216,
            0.3194077597645701,
            0.32076899287451444
          ],
          "mean": 0.31977228384654893,
          "std": 0.000713200415662328
        },
        "rmse_lr_mm": {
          "per_seed": [
            0.5977371082773077,
            0.5968778124758926,
            0.599587242762972
          ],
          "mean": 0.5980673878387242,
          "std": 0.001130506246968718
        },
        "log10_cond_obs": {
          "per_seed": [
            0.3813776742534818,
            0.3813776742534818,
            0.3813776742534818
          ],
          "mean": 0.3813776742534818,
          "std": 0.0
        },
        "n_observations": {
          "per_seed": [
            1160.4444444444443,
            1160.4444444444443,
            1160.4444444444443
          ],
          "mean": 1160.4444444444443,
          "std": 0.0
        },
        "roi_centroid_err_mm__liver_dome": {
          "per_seed": [
            0.4350874037858223,
            0.4340896164057463,
            0.4342536741033093
          ],
          "mean": 0.43447689809829254,
          "std": 0.0004368574351066718
        },
        "roi_centroid_err_mm__liver_inferior": {
          "per_seed": [
            0.3925996926503466,
            0.3940769159994351,
            0.3981389762175021
          ],
          "mean": 0.39493852828909465,
          "std": 0.002342035782904684
        },
        "roi_centroid_err_mm__liver_lateral": {
          "per_seed": [
            0.611596076873094,
            0.6143529977557939,
            0.6127585219151861
          ],
          "mean": 0.612902532181358,
          "std": 0.001130105424562627
        }
      }
    },
    {
      "name": "parallel-2sag",
      "is_baseline": false,
      "what": "2 个矢状切片,LR方向间距 d=50mm(4档中RMSE最优,PASS1池化选出),对称置于ROI中心两侧。",
      "metrics": {
        "dvf_rmse3d_mm": {
          "per_seed": [
            0.7195830749024339,
            0.7187731113259338,
            0.7188964232509512
          ],
          "mean": 0.719084203159773,
          "std": 0.0003563296408589919
        },
        "dvf_p95_mm": {
          "per_seed": [
            1.4493165011608533,
            1.4463285390041793,
            1.4460825602101255
          ],
          "mean": 1.447242533458386,
          "std": 0.0014699507735383623
        },
        "rmse_ap_mm": {
          "per_seed": [
            0.3181534717076165,
            0.3174655947601542,
            0.3173520059731617
          ],
          "mean": 0.31765702414697744,
          "std": 0.0003540910791316606
        },
        "rmse_si_mm": {
          "per_seed": [
            0.3108746013899322,
            0.312141759777104,
            0.31207589865340374
          ],
          "mean": 0.3116974199401467,
          "std": 0.0005824415260129845
        },
        "rmse_lr_mm": {
          "per_seed": [
            0.5656281042673564,
            0.5642853032925785,
            0.5645426513094396
          ],
          "mean": 0.5648186862897915,
          "std": 0.0005819078376338405
        },
        "log10_cond_obs": {
          "per_seed": [
            0.18921735611922205,
            0.18921735611922205,
            0.18921735611922205
          ],
          "mean": 0.18921735611922205,
          "std": 0.0
        },
        "n_observations": {
          "per_seed": [
            2071.1111111111113,
            2071.1111111111113,
            2071.1111111111113
          ],
          "mean": 2071.1111111111113,
          "std": 0.0
        },
        "roi_centroid_err_mm__liver_dome": {
          "per_seed": [
            0.46949751038719967,
            0.47051773549360587,
            0.47083586230950536
          ],
          "mean": 0.47028370273010367,
          "std": 0.0005708910974836945
        },
        "roi_centroid_err_mm__liver_inferior": {
          "per_seed": [
            0.36450324827769676,
            0.36527512238065274,
            0.3626674846500416
          ],
          "mean": 0.3641486184361304,
          "std": 0.0010936987327180893
        },
        "roi_centroid_err_mm__liver_lateral": {
          "per_seed": [
            0.5209300053740122,
            0.5174691810876291,
            0.5200792125973565
          ],
          "mean": 0.5194927996863327,
          "std": 0.0014724664709179869
        }
      }
    },
    {
      "name": "parallel-2cor",
      "is_baseline": false,
      "what": "2 个冠状切片,AP方向间距 d=50mm(同法选出),对称置于ROI中心两侧,alpha=1.0噪声比。",
      "metrics": {
        "dvf_rmse3d_mm": {
          "per_seed": [
            0.7249669898314349,
            0.7260341145674526,
            0.7256056150960917
          ],
          "mean": 0.7255355731649931,
          "std": 0.0004384580592935581
        },
        "dvf_p95_mm": {
          "per_seed": [
            1.4603179614797392,
            1.4634485678457994,
            1.462991861156823
          ],
          "mean": 1.462252796827454,
          "std": 0.0013807814495213264
        },
        "rmse_ap_mm": {
          "per_seed": [
            0.3459543801899958,
            0.34799063475409703,
            0.34657808261388345
          ],
          "mean": 0.3468410325193254,
          "std": 0.0008518372896411323
        },
        "rmse_si_mm": {
          "per_seed": [
            0.32389341289590273,
            0.32318297362427933,
            0.32253470492639397
          ],
          "mean": 0.3232036971488587,
          "std": 0.0005548837325391697
        },
        "rmse_lr_mm": {
          "per_seed": [
            0.5486216913823696,
            0.5491637453424559,
            0.5498713535261732
          ],
          "mean": 0.5492189300836663,
          "std": 0.0005116625744640717
        },
        "log10_cond_obs": {
          "per_seed": [
            0.4506764139687633,
            0.4506764139687633,
            0.4506764139687633
          ],
          "mean": 0.4506764139687633,
          "std": 0.0
        },
        "n_observations": {
          "per_seed": [
            1619.111111111111,
            1619.111111111111,
            1619.111111111111
          ],
          "mean": 1619.111111111111,
          "std": 0.0
        },
        "roi_centroid_err_mm__liver_dome": {
          "per_seed": [
            0.4435227933667892,
            0.44523714944271786,
            0.4431147236857297
          ],
          "mean": 0.4439582221650789,
          "std": 0.0009195547662743229
        },
        "roi_centroid_err_mm__liver_inferior": {
          "per_seed": [
            0.40095578059569664,
            0.40426221725139583,
            0.4037708559983107
          ],
          "mean": 0.4029962846151344,
          "std": 0.001456731835368598
        },
        "roi_centroid_err_mm__liver_lateral": {
          "per_seed": [
            0.5851687101462143,
            0.5888966457607234,
            0.5882380645595413
          ],
          "mean": 0.5874344734888264,
          "std": 0.0016245399232153904
        }
      }
    },
    {
      "name": "orthogonal-sag-cor",
      "is_baseline": false,
      "what": "1 矢状 + 1 冠状,交线过ROI中心;矢状提供(AP,SI),冠状提供(LR,SI),LR方向可被冠状臂观测到。",
      "metrics": {
        "dvf_rmse3d_mm": {
          "per_seed": [
            0.7356201106203191,
            0.7362641159103168,
            0.7362820963088867
          ],
          "mean": 0.7360554409465075,
          "std": 0.00030791253450474914
        },
        "dvf_p95_mm": {
          "per_seed": [
            1.496559130955875,
            1.4985341536778873,
            1.4990447136163998
          ],
          "mean": 1.4980459994167206,
          "std": 0.0010718368120806205
        },
        "rmse_ap_mm": {
          "per_seed": [
            0.33361109887559176,
            0.33469161066073777,
            0.3342979004913908
          ],
          "mean": 0.33420020334257344,
          "std": 0.00044649373203897247
        },
        "rmse_si_mm": {
          "per_seed": [
            0.30805878832036726,
            0.3071214190887099,
            0.3073614123978163
          ],
          "mean": 0.30751387326896445,
          "std": 0.0003975747368214239
        },
        "rmse_lr_mm": {
          "per_seed": [
            0.5787403258756537,
            0.5794331782591143,
            0.5795560380505724
          ],
          "mean": 0.5792431807284468,
          "std": 0.0003590922669941229
        },
        "log10_cond_obs": {
          "per_seed": [
            0.29235378552260516,
            0.29235378552260516,
            0.29235378552260516
          ],
          "mean": 0.29235378552260516,
          "std": 0.0
        },
        "n_observations": {
          "per_seed": [
            2092.6666666666665,
            2092.6666666666665,
            2092.6666666666665
          ],
          "mean": 2092.6666666666665,
          "std": 0.0
        },
        "roi_centroid_err_mm__liver_dome": {
          "per_seed": [
            0.4421104699466801,
            0.4432015066867696,
            0.4438500366247127
          ],
          "mean": 0.4430540044193874,
          "std": 0.0007177932599258657
        },
        "roi_centroid_err_mm__liver_inferior": {
          "per_seed": [
            0.371482306753809,
            0.3743936141013589,
            0.3730760072100195
          ],
          "mean": 0.3729839760217291,
          "std": 0.0011903164628753484
        },
        "roi_centroid_err_mm__liver_lateral": {
          "per_seed": [
            0.574512245134014,
            0.5747744818553259,
            0.5749056895999731
          ],
          "mean": 0.574730805529771,
          "std": 0.00016356518172469075
        }
      }
    },
    {
      "name": "oracle-full-dvf",
      "is_baseline": false,
      "what": "全部网格点三维真值位移直接拟合PCA系数(S=I,无噪声,无切片采样),上界臂,只含K截断误差。",
      "metrics": {
        "dvf_rmse3d_mm": {
          "per_seed": [
            0.7000464521775355,
            0.7000464521775355,
            0.7000464521775355
          ],
          "mean": 0.7000464521775355,
          "std": 0.0
        },
        "dvf_p95_mm": {
          "per_seed": [
            1.4074036872834381,
            1.4074036872834381,
            1.4074036872834381
          ],
          "mean": 1.4074036872834383,
          "std": 2.220446049250313e-16
        },
        "rmse_ap_mm": {
          "per_seed": [
            0.31205099520773993,
            0.31205099520773993,
            0.31205099520773993
          ],
          "mean": 0.31205099520773993,
          "std": 0.0
        },
        "rmse_si_mm": {
          "per_seed": [
            0.33609084685177887,
            0.33609084685177887,
            0.33609084685177887
          ],
          "mean": 0.33609084685177887,
          "std": 0.0
        },
        "rmse_lr_mm": {
          "per_seed": [
            0.5288971112217079,
            0.5288971112217079,
            0.5288971112217079
          ],
          "mean": 0.5288971112217079,
          "std": 0.0
        },
        "log10_cond_obs": {
          "per_seed": [
            6.535997273305607e-16,
            6.535997273305607e-16,
            6.535997273305607e-16
          ],
          "mean": 6.535997273305607e-16,
          "std": 0.0
        },
        "n_observations": {
          "per_seed": [
            27833.666666666668,
            27833.666666666668,
            27833.666666666668
          ],
          "mean": 27833.666666666668,
          "std": 0.0
        },
        "roi_centroid_err_mm__liver_dome": {
          "per_seed": [
            0.4724243113108483,
            0.4724243113108483,
            0.4724243113108483
          ],
          "mean": 0.4724243113108482,
          "std": 5.551115123125783e-17
        },
        "roi_centroid_err_mm__liver_inferior": {
          "per_seed": [
            0.3594989905143546,
            0.3594989905143546,
            0.3594989905143546
          ],
          "mean": 0.3594989905143546,
          "std": 0.0
        },
        "roi_centroid_err_mm__liver_lateral": {
          "per_seed": [
            0.5057445283561814,
            0.5057445283561814,
            0.5057445283561814
          ],
          "mean": 0.5057445283561814,
          "std": 0.0
        }
      }
    }
  ],
  "headline": {
    "metric": "dvf_rmse3d_mm",
    "baseline_mean": 0.719084203159773,
    "proposed_mean": 0.7360554409465075,
    "delta": -0.009215406483709195,
    "claim": "等2切片预算下,正交(矢状+冠状)与最佳间距平行双矢状(d=50mm)在整体三维DVF重建RMSE上差异很小(均值差-0.009mm,-1.1%,27配对样本11/27方向为正,bootstrap95%CI=[-0.017,-0.003]mm);平行矢状臂对LR(穿层)方向的重建误差是正交臂的0.98倍,没有出现预期的穿层误差放大,且LR分量反而占正交臂总平方误差的61.9%,说明瓶颈主要在K=3低秩模型截断而非观测几何——预测被推翻。"
  },
  "deviations": [
    "K/sigma_sag/alpha/interleave/roi_site 五维完整网格未穷举(算力充裕但30分钟硬预算下优先保证核心预测的统计功效);仅对 parallel_spacing_mm 做了4档穷举池化选优(PASS1),其余维度固定在计划给出的默认值(K=3, sigma_sag=0.5mm, alpha=1.0, interleave=false, roi_site=liver_dome 作为headline/ROI主指标,另两个roi_site liver_inferior/liver_lateral 已计算并落入 arms[].metrics 供参考)。",
    "interleave=true(双切片分时交替采集)未实现,因需要额外的时间插值建模,超出30分钟预算范围;本轮结论仅适用于双切片同帧并发观测(interleave=false)的理想化场景。",
    "parallel-2cor 臂的间距选择独立于 parallel-2sag,用相同的池化RMSE最优化流程单独选出,两者最终都选中同一档 d=50mm(4档中的最大值),说明在本数据集网格范围内,越大的切片间距对该低秩模型越有利(见 figs/spacing_sweep.png);这提示50mm可能非真正最优点而是所测网格范围的边界解,未外推到>50mm间距。"
  ],
  "runtime_sec": 6
}