miccai26-code

Codebase for the MICCAI 2026 paper "CDPM-Align: Multi-Scale Guidance-Aligned Diffusion Pretraining for Robust Few-Shot Anatomical Landmark Detection".

Overview

We compare pretraining strategies for few-shot anatomical landmark detection across three public benchmarks:

Dataset	Landmarks	Metric units	Default shots
Shenzhen (chest)	6	pixels	10, 25
ISBI2015 (cephalo)	19	mm	10, 25
DHA (hand)	37	mm	10, 25

Method	Paper name	Config prefix
CDPM-Align (ours)	CDPM-align (λ=5)	`cdpm_align`
CDPM — combined pretraining	CDPM (combined, λ=0)	`cdpm`
CDPM — NIH pretraining	CDPM (NIH, λ=0)	`cdpm_nih`
DINO SSL	ResNet-101 (DINO)	`ssl_dino`
MoCo v3 SSL	ResNet-101 (MoCo v3)	`ssl_mocov3`
SimCLR v2 SSL	ResNet-101 (SimCLR v2)	`ssl_simclrv2`
ImageNet init	ResNet-101 (ImageNet)	`imagenet`
DDPM DiVia	Di Via et al.	`ddpm_divia`
Random init	CDPM Scratch	`random`

Pretraining

Phase 1 — CDPM base (combined chest + hand + cephalo, 45k iterations)

python -m pretraining.cdpm.run -c configs/pretraining/cdpm_combined.json

Phase 1 — CDPM base (NIH ChestX-ray14, 50k iterations)

python -m pretraining.cdpm.run -c configs/pretraining/cdpm_nih.json

Phase 2 — CDPM-Align fine-tuning (5k iterations, λ=5)

python -m pretraining.cdpm_align.run_finetune \
    -c configs/pretraining/cdpm_align.json \
    --lambda_align 5.0 \
    --pretrained_checkpoint checkpoints/pretraining/cdpm/combined_45k.pt

SSL — DINO

python -m pretraining.ssl.ssl_main -c configs/pretraining/ssl_dino.json

SSL — MoCo v3

python -m pretraining.ssl.ssl_main -c configs/pretraining/ssl_mocov3.json

SSL — SimCLR v2

python -m pretraining.ssl.ssl_main -c configs/pretraining/ssl_simclrv2.json

DDPM DiVia (domain-specific, one model per dataset)

# Chest (Shenzhen)
python -m pretraining.ddpm_divia.ddpm_main -c configs/pretraining/ddpm_config.json

# Hand (DHA)
python -m pretraining.ddpm_divia.ddpm_main -c configs/pretraining/ddpm_config.json \
    --dataset hand

# Cephalo (ISBI2015)
python -m pretraining.ddpm_divia.ddpm_main -c configs/pretraining/ddpm_config.json \
    --dataset cephalo

Landmark Detection

All downstream experiments use the NLL loss, AdamW (lr=1e-4), and early stopping (patience=15). Default: 10 training samples. Override with -n N (e.g., -n 25).

Shenzhen (chest)

# CDPM-Align (ours) — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/chest/cdpm_align.json

# CDPM-Align — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/chest/cdpm_align.json -n 25

# CDPM (NIH, λ=0) — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/chest/cdpm_nih.json

# CDPM (NIH, λ=0) — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/chest/cdpm_nih.json -n 25

# CDPM (combined, λ=0) — 10-shot  [ablation: Table 4]
python -m landmark_detection.landmarks_main -c configs/landmarks/chest/cdpm.json

# CDPM (combined, λ=0) — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/chest/cdpm.json -n 25

# Di Via et al. (DDPM) — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/chest/ddpm_divia.json

# Di Via et al. (DDPM) — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/chest/ddpm_divia.json -n 25

# ResNet-101 ImageNet — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/chest/imagenet.json

# ResNet-101 ImageNet — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/chest/imagenet.json -n 25

# ResNet-101 DINO — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/chest/ssl_dino.json

# ResNet-101 DINO — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/chest/ssl_dino.json -n 25

# ResNet-101 MoCo v3 — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/chest/ssl_mocov3.json

# ResNet-101 MoCo v3 — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/chest/ssl_mocov3.json -n 25

# ResNet-101 SimCLR v2 — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/chest/ssl_simclrv2.json

# ResNet-101 SimCLR v2 — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/chest/ssl_simclrv2.json -n 25

# Random init — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/chest/random.json

# Random init — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/chest/random.json -n 25

ISBI2015 (cephalo)

# CDPM-Align (ours) — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/cephalo/cdpm_align.json

# CDPM-Align — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/cephalo/cdpm_align.json -n 25

# CDPM (NIH, λ=0) — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/cephalo/cdpm_nih.json

# CDPM (NIH, λ=0) — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/cephalo/cdpm_nih.json -n 25

# CDPM (combined, λ=0) — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/cephalo/cdpm.json

# CDPM (combined, λ=0) — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/cephalo/cdpm.json -n 25

# Di Via et al. (DDPM) — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/cephalo/ddpm_divia.json

# Di Via et al. (DDPM) — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/cephalo/ddpm_divia.json -n 25

# ResNet-101 ImageNet — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/cephalo/imagenet.json

# ResNet-101 ImageNet — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/cephalo/imagenet.json -n 25

# ResNet-101 DINO — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/cephalo/ssl_dino.json

# ResNet-101 DINO — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/cephalo/ssl_dino.json -n 25

# ResNet-101 MoCo v3 — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/cephalo/ssl_mocov3.json

# ResNet-101 MoCo v3 — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/cephalo/ssl_mocov3.json -n 25

# ResNet-101 SimCLR v2 — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/cephalo/ssl_simclrv2.json

# ResNet-101 SimCLR v2 — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/cephalo/ssl_simclrv2.json -n 25

# Random init — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/cephalo/random.json

# Random init — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/cephalo/random.json -n 25

DHA (hand)

# CDPM-Align (ours) — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/hand/cdpm_align.json

# CDPM-Align — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/hand/cdpm_align.json -n 25

# CDPM (NIH, λ=0) — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/hand/cdpm_nih.json

# CDPM (NIH, λ=0) — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/hand/cdpm_nih.json -n 25

# CDPM (combined, λ=0) — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/hand/cdpm.json

# CDPM (combined, λ=0) — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/hand/cdpm.json -n 25

# Di Via et al. (DDPM) — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/hand/ddpm_divia.json

# Di Via et al. (DDPM) — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/hand/ddpm_divia.json -n 25

# ResNet-101 ImageNet — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/hand/imagenet.json

# ResNet-101 ImageNet — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/hand/imagenet.json -n 25

# ResNet-101 DINO — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/hand/ssl_dino.json

# ResNet-101 DINO — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/hand/ssl_dino.json -n 25

# ResNet-101 MoCo v3 — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/hand/ssl_mocov3.json

# ResNet-101 MoCo v3 — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/hand/ssl_mocov3.json -n 25

# ResNet-101 SimCLR v2 — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/hand/ssl_simclrv2.json

# ResNet-101 SimCLR v2 — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/hand/ssl_simclrv2.json -n 25

# Random init — 10-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/hand/random.json

# Random init — 25-shot
python -m landmark_detection.landmarks_main -c configs/landmarks/hand/random.json -n 25

Checkpoints

Best experiment checkpoints are symlinked under checkpoints/pretraining/:

checkpoints/pretraining/
  cdpm/combined_45k.pt          — CDPM base, 45k iterations, combined dataset (chest+hand+cephalo)
  cdpm/nih_50k.pt               — CDPM base, 50k iterations, NIH ChestX-ray14
  cdpm_align/lambda5.pt         — CDPM-Align, λ=5, best ablation (converted, Config-free)
  ssl/dino/encoder_only.pth     — DINO ResNet-101 encoder
  ssl/mocov3/encoder_only.pth   — MoCo v3 ResNet-101 encoder
  ssl/simclrv2/encoder_only.pth — SimCLR v2 ResNet-101 encoder
  ddpm_divia/chest.pt           — DiVia DDPM, Shenzhen domain
  ddpm_divia/hand.pt            — DiVia DDPM, DHA domain
  ddpm_divia/cephalo.pt         — DiVia DDPM, ISBI2015 domain

Datasets

Data root: /nethome/home/user/Datasets/Landmarks_datasets/

Dataset	Paper name	Landmarks	Default sigma
chest	Shenzhen	6	5
hand	DHA	37	5
cephalo	ISBI2015	19	1

Configs

configs/
  pretraining/
    cdpm_combined.json       — CDPM Phase 1, combined dataset
    cdpm_nih.json            — CDPM Phase 1, NIH ChestX-ray14
    cdpm_align.json          — CDPM-Align Phase 2 fine-tuning
    ssl_dino.json            — DINO SSL
    ssl_mocov3.json          — MoCo v3 SSL
    ssl_simclrv2.json        — SimCLR v2 SSL
    ddpm_config.json         — DDPM DiVia (dataset overridden per run)
  landmarks/
    chest/     cdpm_align.json  cdpm.json  cdpm_nih.json  ssl_dino.json
               ssl_mocov3.json  ssl_simclrv2.json  imagenet.json
               ddpm_divia.json  random.json
    hand/      (same 9 files)
    cephalo/   (same 9 files)

Requirements

pip install -r requirements.txt

Name		Name	Last commit message	Last commit date
Latest commit History 2 Commits
configs		configs
datasets		datasets
landmark_detection		landmark_detection
pretraining		pretraining
utils		utils
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
__init__.py		__init__.py
requirements.txt		requirements.txt

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

miccai26-code

Overview

Pretraining

Phase 1 — CDPM base (combined chest + hand + cephalo, 45k iterations)

Phase 1 — CDPM base (NIH ChestX-ray14, 50k iterations)

Phase 2 — CDPM-Align fine-tuning (5k iterations, λ=5)

SSL — DINO

SSL — MoCo v3

SSL — SimCLR v2

DDPM DiVia (domain-specific, one model per dataset)

Landmark Detection

Shenzhen (chest)

ISBI2015 (cephalo)

DHA (hand)

Checkpoints

Datasets

Configs

Requirements

About

Uh oh!

Releases

Packages

Uh oh!

Contributors

Uh oh!

Languages

Folders and files

Latest commit

History

Repository files navigation

miccai26-code

Overview

Pretraining

Phase 1 — CDPM base (combined chest + hand + cephalo, 45k iterations)

Phase 1 — CDPM base (NIH ChestX-ray14, 50k iterations)

Phase 2 — CDPM-Align fine-tuning (5k iterations, λ=5)

SSL — DINO

SSL — MoCo v3

SSL — SimCLR v2

DDPM DiVia (domain-specific, one model per dataset)

Landmark Detection

Shenzhen (chest)

ISBI2015 (cephalo)

DHA (hand)

Checkpoints

Datasets

Configs

Requirements

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages