| { |
| "model_type": "ResEncLUNet", |
| "task": "binary 3D segmentation of amyloid-beta plaques in light-sheet microscopy (all LANTERN Abeta cohorts)", |
| "framework": "pytorch", |
| "num_input_channels": 1, |
| "num_classes": 2, |
| "encoder": { |
| "name": "ResEncL", |
| "n_stages": 6, |
| "features_per_stage": [ |
| 32, |
| 64, |
| 128, |
| 256, |
| 320, |
| 320 |
| ], |
| "frozen": true, |
| "init": "nnBYOL3D self-supervised pretraining on ~50k near-iso-4 light-sheet patches" |
| }, |
| "params_total": 102462826, |
| "params_trainable": 12166250, |
| "input": { |
| "patch_size": [ |
| 128, |
| 128, |
| 128 |
| ], |
| "axes_order": "ZYX", |
| "voxel_size_um": { |
| "z": "2.75-4.4 (cohort dependent)", |
| "y": 3.25, |
| "x": 3.25 |
| }, |
| "dtype": "float32", |
| "note": "raw intensities; NOT N4-corrected or rescaled" |
| }, |
| "preprocessing": { |
| "method": "per-tile percentile clip then z-score", |
| "clip_percentiles": [ |
| 0.5, |
| 99.5 |
| ], |
| "order": "clip first, then mean/std computed on the clipped volume" |
| }, |
| "inference": { |
| "n_folds": 5, |
| "combine": "sum of per-fold plaque decisions (softmax[1] >= 0.5); plaque where votes >= 3", |
| "prob_threshold": 0.5, |
| "tiling_stride": 64, |
| "tile_overlap_combine": "mean of per-fold softmax over overlapping tiles (not max)", |
| "recommended_precision": "bf16", |
| "recommended_batch_size": 8, |
| "benchmark_ms_per_tile_L40S": 27.8, |
| "recommended_postprocess": "drop predicted plaque components < 8 near-iso voxels" |
| }, |
| "training": { |
| "n_patches": 178, |
| "n_subjects": 68, |
| "set": "v7/total_set/v1", |
| "composition": { |
| "ki3": 74, |
| "o22b": 4, |
| "mc4": 53, |
| "prune": 4, |
| "vac": 3, |
| "mapt": 40 |
| }, |
| "empty_plaque_patches": 19, |
| "patches_with_label2_folded_to_background": 25, |
| "cv": "5-fold, random patch-level split (NOT subject-grouped)", |
| "split_seed": 1234, |
| "gt_min_connected_component_voxels": 4, |
| "label_mode": "plaque", |
| "loss": "DiceCE (batch-Dice on plaque)", |
| "decoder_init": "random (no warm start)", |
| "encoder": "frozen SSL ResEncL (byol_c128)", |
| "epochs": 250, |
| "optimizer": "AdamW", |
| "lr": 0.001, |
| "warmup_epochs": 10, |
| "schedule": "cosine", |
| "batch_size": 2, |
| "accumulate_grad_batches": 4, |
| "precision": "bf16-mixed", |
| "checkpoint": "best-val object F1" |
| }, |
| "evaluation": { |
| "primary_metric": "object F1 @ IoU>=0.5, 26-connectivity, per-fold validation split (36 patches), mean over 5 folds at each fold's best epoch", |
| "object_f1": 0.4589, |
| "object_f1_sd_over_folds": 0.06, |
| "precision": 0.426, |
| "recall": 0.501, |
| "fp_per_patch": 100.7, |
| "object_f1_with_pred_min_cc_8": 0.5042, |
| "fp_per_patch_with_pred_min_cc_8": 70.0, |
| "voxel_dice": 0.605, |
| "per_fold": [ |
| { |
| "fp_per_patch": 108.88888549804688, |
| "fp_per_patch_predfilt": 74.02777862548828, |
| "loss": 0.3341991901397705, |
| "object_f1": 0.4092409312725067, |
| "object_f1_predfilt": 0.4598941504955292, |
| "object_precision": 0.36280885338783264, |
| "object_recall": 0.4693019390106201, |
| "voxel_dice": 0.6610821485519409, |
| "epoch": 99 |
| }, |
| { |
| "fp_per_patch": 81.22222137451172, |
| "fp_per_patch_predfilt": 58.83333206176758, |
| "loss": 0.3775239586830139, |
| "object_f1": 0.4456420838832855, |
| "object_f1_predfilt": 0.4795815646648407, |
| "object_precision": 0.45058247447013855, |
| "object_recall": 0.44080883264541626, |
| "voxel_dice": 0.7066648602485657, |
| "epoch": 64 |
| }, |
| { |
| "fp_per_patch": 99.97222137451172, |
| "fp_per_patch_predfilt": 66.44444274902344, |
| "loss": 0.2981194853782654, |
| "object_f1": 0.5641064643859863, |
| "object_f1_predfilt": 0.6198633313179016, |
| "object_precision": 0.5060389637947083, |
| "object_recall": 0.6372277736663818, |
| "voxel_dice": 0.6496661901473999, |
| "epoch": 119 |
| }, |
| { |
| "fp_per_patch": 122.68571472167969, |
| "fp_per_patch_predfilt": 88.71428680419922, |
| "loss": 0.49983668327331543, |
| "object_f1": 0.397587388753891, |
| "object_f1_predfilt": 0.4390422999858856, |
| "object_precision": 0.3591044843196869, |
| "object_recall": 0.4453081488609314, |
| "voxel_dice": 0.5200831294059753, |
| "epoch": 109 |
| }, |
| { |
| "fp_per_patch": 90.97142791748047, |
| "fp_per_patch_predfilt": 61.88571548461914, |
| "loss": 0.5084365010261536, |
| "object_f1": 0.47774097323417664, |
| "object_f1_predfilt": 0.5225977301597595, |
| "object_precision": 0.44923022389411926, |
| "object_recall": 0.5101158618927002, |
| "voxel_dice": 0.4864481985569, |
| "epoch": 24 |
| } |
| ], |
| "note": "patch-level CV with same-subject leakage; validation metric at the selected (best) epoch, so optimistic; LR 3e-4/1e-4 within 0.02; GT floor 8 variant +0.02 raw F1 / +0.05 pred-filtered F1 at identical voxel Dice" |
| }, |
| "license": "cc-by-nc-4.0", |
| "code": "https://github.com/Arshya-Guru/LANTERN", |
| "classes": [ |
| "background", |
| "plaque" |
| ], |
| "hub_repo": "apooladi/lantern-abeta", |
| "arm": "scratchdec_mincc4_cond", |
| "weights": "best-val", |
| "bottleneck_conditioner": { |
| "enabled": true, |
| "inputs": [ |
| "log1p(mean of raw tile)", |
| "log1p(std of raw tile)" |
| ], |
| "where": "bottleneck (deepest encoder features), Linear(2->32) broadcast + 1x1x1 conv mix", |
| "standardisation": "per-fold buffers in state_dict (cond.log_mean_mu, cond.log_mean_sd, cond.log_std_mu, cond.log_std_sd)", |
| "train_dropout": 0.3, |
| "required_at_inference": true, |
| "forward_signature": "net(x, raw_mean=torch.tensor([[m]]), raw_std=torch.tensor([[s]])) # m, s over the RAW (unnormalised) 128^3 tile" |
| } |
| } |
