These are used to evaluate FeatDice scores mentioned in our paper.
This is used to train and test PlaneCycle on different configurations.
The backbone is built according to backbone_model.block_type in the cfg: PlaneCycle loads a plain 2D backbone via torch.hub and converts it with planecycle_converter; Slice2D / Flatten3D use BaselineViT from experiments/baselines/dinov3/vit (its state-dict keys equal the plain backbone's, so the official checkpoints load unchanged).
This folder contains full fine-tuning and linear probing example files of configuration.
Notably, block_type sits at the backbone_model level (not inside load_model — the hub entrypoints no longer accept it), and load_model.repo_or_dir should point to the repository root (where hubconf.py lives).
MultiSegmentationLoss provides a combined loss that can select a combination from Cross-Entropy Loss, Dice Loss and Boundary Dice Loss. Notably, all losses are computed as volumetric loss.
WarmupOneCycleLR realizes linear warmup with cosine annealing following.
build_scheduler contains various choices for schedulers, including PyTorch built-in schedulers and WarmupOnceCycleLR.
Notably, this file is edited basing on an original DINOv3 version.
_get_backbone_out_indices can select to concatenate the number of layers of outputs from a backbone，selecting from "last", "four_last", "four_even_intervals".
ModelWithIntermediateLayers pack up the backbone into a new class，which allows to straightly use get_intermediate_layers to extract the outputs from the backbone.
ProgressiveUpHead is a Progressively Upsampling 3D Decoder.
UpBlock3D is the smallest upsampling block inProgressiveUpHead.
FeatureDecoder is a wrapper combining with a backbone and a head.
build_segmentation_decoderis a function which will return a FeatureDecoder type model.
LIDCSegDataset is the dataset structure specified for LIDC, which returns transformed data and gt-mask as tensors.
Transform contains fundamental preprocessing and augmentation for LIDC, which is used in LIDCSegDataset.
Notably, for the gt-mask, only voxels gaining at least half votes from the experts will be marked as lesions.
The file contains necessary transforms for the LIDC dataset, including rotation, crop and random_center.
MMWHS_Dataset is the main implementation for MMWHS data, which contains augmentations and preprocessing.
Notably, the original data also contains SDF etc. but we only adopt segmentation masks as gt for efficiency.
cd PlaneCycle # Project Root Directory
# Initialize configuration files basing on needs, there are examples for considerations.
python3 -m experiments.segmentation.seg_training --cfg "*.yaml"