# ===================== ToothCanal-SDF default config =====================
paths:
  # On AutoDL, the large data disk is /root/autodl-tmp (system disk /root is small).
  hf_repo: "zgy0823/label"
  raw_dir: "/root/autodl-tmp/toothcanal/data/raw"
  proc_dir: "/root/autodl-tmp/toothcanal/data/processed"
  out_dir: "/root/autodl-tmp/toothcanal/outputs"
  # Drop the supplementary 21-30 TOOTH-label zip's contents here (see run notes).
  # Files named by case number, e.g. 21.nii.gz ... 30.nii.gz  (or a per-case folder).
  extra_tooth_dir: "/root/autodl-tmp/toothcanal/extra_tooth_labels"
  # Zips whose name contains any of these substrings are NOT downloaded/used.
  skip_zip_substrings: ["36-40"]

label_scheme:
  n_teeth: 28                        # 28 tooth positions
  canal_lo: 1                        # canal/pulp labels: 1..28  (inner)
  canal_hi: 28
  body_lo: 29                        # tooth-body labels: 29..56 (outer)
  body_hi: 56
  pair_offset: 28                    # instance i: canal=label i, body=label i+28
  pair_mode: geometric               # geometric = assign each canal to the body it sits inside
                                     #   (robust to mis-numbered annotations; fixes 032/035/013/014).
                                     #   set to 'offset' to restore legacy canal=i / body=i+28 pairing.

preprocess:
  spacing: [0.4, 0.4, 0.4]           # mm; whole-volume working resolution (memory friendly)
  clip_hu: [-1000, 3000]             # intensity clip before normalization
  descriptor_erode_iter: 3           # erosion iters to build separable tooth-core descriptors
  # ---- canal cleaning ----
  speck_min_voxels: 30               # remove connected components smaller than this (keeps real multi-canals)
  closing_radius: 0                  # 0 = no morphological closing (avoids canal dilation at 0.25mm).
                                     # set to 1 only if your canals have many artifact-induced breaks.
  fill_holes: true

split:
  # Cases 1-30 -> training (5-fold CV).  Cases 31-35 -> held-out eval.
  # 36-40 has no usable data and is excluded.
  train_range: [1, 30]
  test_range: [31, 35]
  exclude_range: [36, 99]
  exclude_cases: [26]                # 026 has tooth labels but NO canal labels (would teach
                                     #   the model "no canal" for 28 teeth). add 19,20 too if
                                     #   you want to drop the near-empty partial annotations.
  n_folds: 5
  seed: 42

stage1:                              # coarse semantic seg: 0 bg / 1 tooth / 2 canal / 3 descriptor-core
  patch_size: [96, 96, 96]
  batch_size: 2
  num_classes: 4
  channels: [16, 32, 64, 128, 256]
  lr: 2.0e-4
  max_epochs: 300
  samples_per_volume: 4
  val_interval: 10
  use_descriptor: true               # use class-3 cores for instance localization at inference
  min_core_voxels: 80                # drop tiny spurious cores (false detections)
  core_merge_mm: 3.0                 # merge over-split cores whose centroids are <this apart
                                     #   (fixes Stage-1 one-tooth-as-many). 0 disables merging.

stage2:                              # per-tooth dual-SDF implicit reconstruction
  roi_mm: 24.0                       # MATCHES the current stage2.pt (trained at 24mm). do NOT raise
  ckpt_name: stage2.pt               # the current good model; retrain configs use a different name
  roi_vox: 96                        # 24/96 = 0.25 mm. matches current stage2.pt
  roi_center: com                    # matches how the current checkpoint was trained
  latent_dim: 64
  enc_channels: [32, 64, 128]       # U-Net encoder (fine+coarse local features)
  feat_dim: 96                       # per-point local feature width (was 64)
  mlp_hidden: 384                    # wider decoder (was 256)
  mlp_layers: 6                      # deeper decoder (was 5)
  encoder_latent: true               # predict latent z = enc(ROI) (generalizes to new cases)
  points_per_tooth: 8192             # more query points (was 6144)
  canal_point_frac: 0.45             # fraction near the canal surface
  centerline_frac: 0.15              # extra fraction near the canal centerline
  canal_weight: 3.0                  # up-weight the small canal (was 2.5)
  near_surface_ratio: 0.65
  near_surface_sigma_mm: 0.5
  sdf_clamp_mm: 2.0
  occ_tau_mm: 0.3
  loss_weights:
    sdf: 1.0
    eikonal: 0.1
    occ: 1.0
    normal: 0.15
    nest: 4.0
    prior: 1.0e-3
    centerline: 0.2                  # softened clDice continuity (was 0.5, caused phantom blobs)
    smoothness: 0.5                  # Duan-2021-style anti-phantom / volume control
  order_weight: 1.0                  # weight of the hard SDF-ordering containment term
  lr: 4.0e-4
  lr_latent: 1.0e-3
  lr_min: 1.0e-5                     # cosine anneal floor
  batch_teeth: 3                     # local-feature grids use more memory
  max_epochs: 700                    # deeper run for the bigger model
  margin_mm: 0.2                     # containment margin
  augment: true                      # random 3D flips + 90-deg rotations (small-data regime)

tto:                                 # test-time optimization of the latent code
  steps: 500                         # was 200 -- helps the long-tail cases (032/035)
  lr: 5.0e-3

infer:
  grid: 128                          # was 96 -- finer marching cubes (smoother thin canals)
  mc_level: -0.2                     # legacy fallback (used if the split knobs below are absent)
  mc_level_tooth: 0.0                # Tier 1: tooth meshed at GT level (0.0) -> removes -0.12 undersize
  mc_level_canal: -0.2              # canal body; scan -0.2/-0.25/-0.3 against signed_rvd (Tier 2)
  tooth_watertight_postprocess: true # tooth: keep largest component + fill holes (good)
  canal_watertight_postprocess: false  # canal: do NOT fill -> preserves apex foramen + fine branches (Tier 2)
  pad_roi: true                      # pad the SDF grid border so MC closes boundary openings
  roi_source: "oracle"               # "oracle"=GT instances (upper bound); "predicted"=Stage-1 drives ROIs
  use_tto: false                     # main results use NO test-time optimization (avoids GT leakage)
  canal_min_component_frac: 0.02     # 0.08->0.02 (Tier 2): keep small apical branches; rely on
                                     # largest (removes SDF bridge stubs; lower it if real thin
                                     # accessory canals are being discarded)

eval:
  n_surface_samples: 30000
  apex_mm: 3.0                        # apical region size (mm) for root-tip-only canal metrics
  gt_wise_predicted: true            # predicted eval: score 1 best prediction per GT tooth
  gt_roi_vox: 144                    # GT meshed at 24/144=0.167mm, FIXED regardless of model,
  gt_grid: 144                       #   so 96- and 144-voxel models are compared on the SAME GT