Alveolar Canal Segmentation

3D Alveolar Canal Segmentation — Multi-Model Benchmark

Pretrained weights for six 3D segmentation models trained on the ToothFairy2 dataset for alveolar canal segmentation from CBCT scans.

Model Performance (Test Set, n=50)

Model	Dice ↑	IoU ↑	PPV ↑	Recall ↑	HD95 (mm) ↓	MaxHD (mm) ↓
nnUNetv2	0.9163 ± 0.0119	0.8458 ± 0.0201	0.8898 ± 0.0326	0.9472 ± 0.0396	1.17 ± 0.31	5.87 ± 3.22
AttentionUnet	0.8895 ± 0.0476	0.8037 ± 0.0659	0.8882 ± 0.0388	0.8980 ± 0.0844	4.93 ± 13.46	10.95 ± 17.03
SwinUNETR	0.8859 ± 0.0495	0.7985 ± 0.0745	0.9021 ± 0.0329	0.8776 ± 0.0949	11.99 ± 22.04	19.97 ± 27.70
SegResNet	0.8760 ± 0.0613	0.7839 ± 0.0843	0.8865 ± 0.0383	0.8755 ± 0.1057	9.71 ± 23.73	17.57 ± 27.08
3D UNet	0.8678 ± 0.0639	0.7713 ± 0.0885	0.8747 ± 0.0432	0.8705 ± 0.1065	12.38 ± 23.20	20.33 ± 27.67
DynUNet	0.7129 ± 0.0515	0.5563 ± 0.0618	0.5607 ± 0.0618	0.9865 ± 0.0262	8.14 ± 8.06	21.33 ± 11.52

All differences between models are statistically significant (Wilcoxon signed-rank, p<0.001) except AttentionUnet vs SwinUNETR (p=0.20, ns).

Repository Structure

abhijithkj3690/alveolar-canal-segmentation/
├── unet/
│   └── best_model.pth           # 3D UNet (MONAI)
├── attention_unet/
│   └── best_model.pth           # AttentionUnet (MONAI)
├── dynunet/
│   └── best_model.pth           # DynUNet (MONAI)
├── swinunetr/
│   └── best_model.pth           # SwinUNETR (MONAI)
├── segresnet/
│   └── best_model.pth           # SegResNet (MONAI)
└── nnunetv2/
    ├── plans.json
    ├── dataset.json
    ├── dataset_fingerprint.json
    └── fold_0/
        ├── checkpoint_best.pth
        └── checkpoint_final.pth

Usage

Installation

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
pip install monai nibabel scipy scikit-image tqdm
pip install huggingface_hub

Load and run inference (MONAI models)

import torch
from monai.networks.nets import AttentionUnet
from huggingface_hub import hf_hub_download

# Download weights
weights_path = hf_hub_download(
    repo_id="abhijithkj3690/alveolar-canal-segmentation",
    filename="attention_unet/best_model.pth",
)

# Build model — must match training architecture exactly
model = AttentionUnet(
    spatial_dims=3,
    in_channels=1,
    out_channels=1,
    channels=(32, 64, 128, 256, 320),
    strides=(2, 2, 2, 2),
    dropout=0.1,
)

model.load_state_dict(
    torch.load(weights_path, map_location="cpu", weights_only=True)
)
model.eval()

Available model strings

Model	filename argument
3D UNet	"unet/best_model.pth"
AttentionUnet	"attention_unet/best_model.pth"
DynUNet	"dynunet/best_model.pth"
SwinUNETR	"swinunetr/best_model.pth"
SegResNet	"segresnet/best_model.pth"

MONAI model constructors

from monai.networks.nets import UNet, AttentionUnet, DynUNet, SwinUNETR, SegResNet

# 3D UNet
UNet(spatial_dims=3, in_channels=1, out_channels=1,
     channels=(32,64,128,256,320), strides=(2,2,2,2), num_res_units=2)

# AttentionUnet
AttentionUnet(spatial_dims=3, in_channels=1, out_channels=1,
              channels=(32,64,128,256,320), strides=(2,2,2,2), dropout=0.1)

# DynUNet
DynUNet(spatial_dims=3, in_channels=1, out_channels=1,
        kernel_size=[[3,3,3]]*5,
        strides=[[1,1,1],[2,2,2],[2,2,2],[2,2,2],[2,2,2]],
        upsample_kernel_size=[[2,2,2]]*4,
        filters=(32,64,128,256,320), dropout=0.1,
        deep_supervision=False, res_block=True)

# SwinUNETR
SwinUNETR(in_channels=1, out_channels=1, feature_size=48,
          depths=(2,2,2,2), num_heads=(3,6,12,24), window_size=7,
          dropout_path_rate=0.1, spatial_dims=3)

# SegResNet
SegResNet(spatial_dims=3, in_channels=1, out_channels=1,
          init_filters=32, blocks_down=(1,2,2,4), blocks_up=(1,1,1),
          dropout_prob=0.1, use_conv_final=True)

Preprocessing

All MONAI models use the following preprocessing (must be applied before inference):

from monai.transforms import Compose, LoadImaged, EnsureChannelFirstd, Orientationd, \
    Spacingd, ScaleIntensityRanged, SpatialPadd, DivisiblePadd, EnsureTyped

transforms = Compose([
    LoadImaged(keys=["image"]),
    EnsureChannelFirstd(keys=["image"]),
    Orientationd(keys=["image"], axcodes="RAS", labels=None),
    Spacingd(keys=["image"], pixdim=(0.3, 0.3, 0.3), mode="bilinear"),
    ScaleIntensityRanged(keys=["image"],
        a_min=-167.8, a_max=1111.5, b_min=0.0, b_max=1.0, clip=True),
    SpatialPadd(keys=["image"], spatial_size=(128, 128, 128)),
    DivisiblePadd(keys=["image"], k=32),
    EnsureTyped(keys=["image"]),
])

Training Details

Setting	Value
Dataset	ToothFairy2 (CBCT, alveolar canal)
Train / Val / Test split	296 / 86 / 50
Patch size	128 × 128 × 128
Target spacing	0.3 × 0.3 × 0.3 mm
Loss	DiceFocalLoss (γ=2.0)
Optimizer	SGD (lr=0.01, momentum=0.99, Nesterov)
LR schedule	Polynomial decay (power=0.9)
Epochs	300 (scratch)
Framework	MONAI 1.3.2 / PyTorch 2.2.2

Code

Full training, inference, and analysis scripts: 👉 github.com/abhijithkj3690/alveolar-canal-segmentation

Dataset

ToothFairy2 Challenge — 3D CBCT scans with expert annotations of the alveolar canal (inferior alveolar nerve canal).

License

MIT — free to use for research and commercial purposes with attribution.