karimii0214/Thyrops_ConvNext-base_RSNA

Thyrops ConvNeXt-Base RSNA

Research checkpoint for pediatric bone-age regression from a grayscale hand radiograph and recorded sex. The model was trained on the RSNA Pediatric Bone Age Challenge training release and returns bone age in years.

This release is for research only. It is not a medical device, is not clinically validated, and must not be used for diagnosis or treatment decisions.

Files

File	Purpose
model.safetensors	Recommended inference-only FP32 model state
training_checkpoint.pth	Original epoch-7 PyTorch checkpoint, including AdamW and scheduler state
modeling_thyrops.py	Self-contained model architecture
preprocessing_thyrops.py	Exact released PNG/JPEG inference preprocessing
inference.py	Single-image command-line inference
config.json	Architecture and preprocessing contract
benchmark_results.json	Machine-readable evaluation summary
checksums.sha256	SHA-256 integrity manifest

Architecture

• ConvNeXt-V2-Base grayscale backbone (timm, 1 input channel)
• Feature Pyramid Network over stages with channels [128, 256, 512, 1024]
• 256-channel aggregation by concatenation, 1x1 convolution, and global average pooling
• sex-conditioned FiLM (female=0, male=1)
• regression head in years and an auxiliary 20-bin ordinal head
• 90,248,330 parameters

The exact tag of the upstream timm pretrained initialization was not recorded at training time. This limits full initialization provenance even though the released final state is deterministic and integrity-checked.

Preprocessing Contract

1. Convert the rendered radiograph to grayscale.
2. Map 8-bit pixels to the project's pseudo-HU convention with pixel * (4000 / 255).
3. Apply window center 600 and window width 2000.
4. Resize with preserved aspect ratio and letterbox to 1024x1024.
5. Pad with 0.523483 and normalize with mean 0.523483, standard deviation 0.284501.
6. Supply sex as 1.0 for male or 0.0 for female.

The pseudo-HU mapping is a project normalization convention for rendered PNG images. It does not recover physical Hounsfield units.

Usage

pip install -r requirements.txt
python inference.py \
  --image hand.png \
  --sex female \
  --weights model.safetensors

Use --bf16 for CUDA BF16 autocast. FP32 is the default and matches the primary benchmark rows.

Validation Results

Results below use fixed 40-image validation subsets from benchmark manifest SHA-256 2f17acf1e1fe93783ac6e3b9521a63c9e9e42b6ca09a2ba008eb3e296d13665f.

Domain	MAE (years)	RMSE (years)	Pearson r	Within 1 year	Within 2 years
RSNA validation	0.3796	0.4975	0.9915	92.5%	100.0%
RHPE domain-shift validation	2.5214	3.0581	0.6020	22.5%	45.0%

These are small validation measurements, not independent clinical test results. The original training split manifest is unavailable, so overlap between the 40-image RSNA diagnostic subset and the original training split cannot be excluded. Official test releases available during this evaluation were unlabeled, so no test MAE is claimed. The RHPE result demonstrates substantial domain-shift degradation.

Precision Sensitivity

The following results are weight quantize-dequantize quality ablations with FP32 GPU execution. They do not represent native BF16/FP8/FP4 latency or memory use.

Weight ablation	RSNA MAE	Delta	RHPE MAE	Delta
FP32 baseline	0.3796	baseline	2.5214	baseline
BF16 QDQ	0.3790	-0.15%	2.5209	-0.02%
FP8 E4M3 QDQ	0.3808	+0.32%	2.5312	+0.39%
NVFP4 E2M1 block-16 QDQ	0.3879	+2.17%	2.5826	+2.43%

Training Checkpoint

training_checkpoint.pth contains:

• model state
• AdamW optimizer state
• CosineAnnealingWarmRestarts scheduler state
• epoch (7)
• best training-time validation MAE (0.6927681565284729 years)
• early-stopping counter

Only load pickle-based PyTorch checkpoints from trusted sources. Use model.safetensors for inference.

Limitations

• The missing original split manifest prevents a strict independent RSNA evaluation claim.
• The compact benchmark subsets are too small to characterize rare pathologies or demographic subgroups.
• Recorded sex is a required model input and may encode dataset-specific bias.
• RHPE evaluation shows poor transfer outside the RSNA domain.
• PNG/JPEG inference is provided; raw DICOM deployment is not validated in this release.
• Distribution shift across scanners, institutions, ages, and acquisition protocols can materially reduce accuracy.
• Predictions require expert review in any research workflow involving patient data.

No training or evaluation images are included in this repository. The source RSNA data terms permit academic research, education, and other non-commercial uses subject to attribution requirements.