A LoRA-fine-tuned MedSigLIP-448 model for pediatric bone age estimation from hand/wrist X-rays. Achieves 8.81-month MAE on the RSNA Bone Age validation set (Pearson r = 0.963), surpassing typical human expert inter-reader variability of 10-13 months.

Part of the Chronos-MSK multi-agent bone age assessment system, built for the Google HAI-DEF competition.

Model Description

This model predicts bone age in months from a single left hand/wrist radiograph. Instead of regressing to a single number, it uses Deep Label Distribution Learning (DLDL) — predicting a full probability distribution over 228 monthly age bins. The final age estimate is the expected value of this distribution, which naturally captures prediction uncertainty.

Biological sex is incorporated as a conditioning signal through a learned embedding, reflecting the known differences in skeletal maturation rates between males and females.

Architecture

Input: Hand/Wrist X-Ray (448×448, letterbox padded)
  │
  ▼
MedSigLIP-448 Vision Encoder (frozen base + DoRA adapters)
  │
  ▼ last_hidden_state (sequence of patch embeddings)
  │
  ▼
Global Average Pooling + LayerNorm → 1152-D visual features
  │
  ├── Sex Input (0/1) → Linear(1,64) → GELU → Linear(64,128) → 128-D
  │
  ▼
Concatenate [visual_1152; sex_128] → 1280-D
  │
  ▼
LayerNorm → Linear(1280,512) → GELU → Dropout(0.1) → Linear(512,228)
  │
  ▼
Softmax → Probability distribution over 228 months
  │
  ▼
Expected Value: â = Σ(k × P(k)) for k = 0..227

Key Design Choices

DoRA (Weight-Decomposed LoRA): rank=16, alpha=32 — adapts the frozen MedSigLIP encoder with only ~2M trainable parameters while preserving medical domain knowledge
DLDL Loss: KL divergence between predicted distribution and a Gaussian target centered on the true age (sigma=12 months) — smoother gradients than MSE regression
Sex Conditioning: Learned embedding concatenated with visual features, enabling sex-specific age estimation without separate models
Test-Time Augmentation: Average predictions from original and horizontally-flipped image for improved accuracy

Results

Evaluated on 1,425 held-out RSNA Bone Age validation cases (strict no-leakage protocol):

Metric	Value
Mean Absolute Error	8.81 months (0.73 years)
Median Absolute Error	7.27 months
Root Mean Square Error	11.39 months
Pearson r	0.963
R²	0.927
Within ±6 months	42.9%
Within ±12 months	73.2%
Within ±24 months	95.7%

Age-Stratified Performance

Age Range	Cases	MAE
0-5 years	94	9.54 months
5-10 years	394	10.04 months
10-15 years	809	8.07 months
15-19 years	124	8.45 months

Sex-Stratified Performance

Sex	Cases	MAE
Male	773	8.26 months
Female	652	9.47 months

Usage

Installation

pip install torch transformers peft numpy opencv-python huggingface_hub

Inference

python

import torch
import torch.nn as nn
import torch.nn.functional as F
import cv2
import json
import numpy as np
from transformers import SiglipVisionModel
from peft import PeftModel
from huggingface_hub import hf_hub_download, snapshot_download


# ── Model Components ──

class GlobalAveragePooling(nn.Module):
    def __init__(self, hidden_size):
        super().__init__()
        self.norm = nn.LayerNorm(hidden_size)

    def forward(self, x):
        return self.norm(x.mean(dim=1))


def letterbox_resize(image, size=448):
    """Resize preserving aspect ratio with zero-padding."""
    h, w = image.shape[:2]
    scale = size / max(h, w)
    nh, nw = int(h * scale), int(w * scale)
    resized = cv2.resize(image, (nw, nh))
    padded = np.zeros((size, size, 3), dtype=np.uint8)
    top, left = (size - nh) // 2, (size - nw) // 2
    padded[top:top + nh, left:left + nw] = resized
    return padded


# ── Load Model ──

def load_regressor(repo_id="rohitrajesh/chronos-msk-regressor", device="cuda"):
    """Load the complete bone age regressor from HuggingFace."""
    
    # Download all files
    config_path = hf_hub_download(repo_id, "config.json")
    heads_path = hf_hub_download(repo_id, "heads.pth")
    adapter_dir = snapshot_download(repo_id, allow_patterns=["adapter/*"])
    
    # Load config
    with open(config_path) as f:
        config = json.load(f)
    
    # Load MedSigLIP base model with LoRA adapter
    base_model = SiglipVisionModel.from_pretrained(
        config["model_id"], torch_dtype=torch.float32
    )
    backbone = PeftModel.from_pretrained(base_model, f"{adapter_dir}/adapter")
    
    hidden = config["hidden_size"]
    num_bins = config["num_bins"]
    
    # Build prediction heads
    pooler = GlobalAveragePooling(hidden)
    gender_embed = nn.Sequential(
        nn.Linear(1, 64), nn.GELU(), nn.Linear(64, 128)
    )
    classifier = nn.Sequential(
        nn.LayerNorm(hidden + 128),
        nn.Linear(hidden + 128, 512),
        nn.GELU(),
        nn.Dropout(0.1),
        nn.Linear(512, num_bins),
    )
    
    # Load trained head weights
    state = torch.load(heads_path, map_location=device)
    pooler.load_state_dict(state["pooler"])
    gender_embed.load_state_dict(state["gender_embed"])
    classifier.load_state_dict(state["classifier"])
    
    # Move to device and set eval mode
    backbone.to(device).eval()
    pooler.to(device).eval()
    gender_embed.to(device).eval()
    classifier.to(device).eval()
    
    return backbone, pooler, gender_embed, classifier, config


# ── Predict ──

def predict_bone_age(image_path, is_male, backbone, pooler,
                     gender_embed, classifier, config, device="cuda"):
    """
    Predict bone age from a hand/wrist X-ray.
    
    Args:
        image_path: Path to X-ray image (any common format)
        is_male: Boolean, biological sex
        
    Returns:
        age_months: Predicted bone age in months (float)
    """
    size = config["image_size"]
    
    # Load and preprocess
    img = cv2.imread(image_path)
    if img is None:
        raise ValueError(f"Cannot read image: {image_path}")
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = letterbox_resize(img, size)
    
    # Normalize: (pixel / 255 - 0.5) / 0.5
    norm = (img.astype(np.float32) / 255.0 - 0.5) / 0.5
    t_img = torch.from_numpy(norm.transpose(2, 0, 1)).unsqueeze(0).to(device)
    t_sex = torch.tensor([[1.0 if is_male else 0.0]], device=device)
    
    with torch.no_grad():
        # Forward pass
        out = backbone(pixel_values=t_img)
        vis = pooler(out.last_hidden_state)
        gen = gender_embed(t_sex)
        logits = classifier(torch.cat([vis, gen], dim=-1))
        
        # Expected value from probability distribution
        probs = F.softmax(logits, dim=-1)
        ages = torch.arange(probs.shape[-1], device=device, dtype=torch.float32)
        predicted_months = torch.sum(probs * ages).item()
    
    return predicted_months


def predict_with_tta(image_path, is_male, backbone, pooler,
                     gender_embed, classifier, config, device="cuda"):
    """
    Predict with Test-Time Augmentation (original + horizontal flip).
    This is the recommended inference method.
    """
    # Original prediction
    pred_orig = predict_bone_age(
        image_path, is_male, backbone, pooler,
        gender_embed, classifier, config, device
    )
    
    # Flipped prediction
    img = cv2.imread(image_path)
    img_flipped = cv2.flip(img, 1)  # Horizontal flip
    tmp_path = "/tmp/chronos_flip.png"
    cv2.imwrite(tmp_path, img_flipped)
    
    pred_flip = predict_bone_age(
        tmp_path, is_male, backbone, pooler,
        gender_embed, classifier, config, device
    )
    
    # Average
    return (pred_orig + pred_flip) / 2.0


# ── Example ──

if __name__ == "__main__":
    device = "cuda" if torch.cuda.is_available() else "cpu"
    print(f"Using device: {device}")
    
    # Load model (downloads from HuggingFace on first run)
    print("Loading model...")
    backbone, pooler, gender_embed, classifier, config = load_regressor(device=device)
    print("Model loaded!")
    
    # Single prediction
    age = predict_bone_age(
        "hand_xray.png",
        is_male=True,
        backbone=backbone,
        pooler=pooler,
        gender_embed=gender_embed,
        classifier=classifier,
        config=config,
        device=device,
    )
    print(f"Predicted bone age: {age:.1f} months ({age/12:.1f} years)")
    
    # With TTA (recommended)
    age_tta = predict_with_tta(
        "hand_xray.png",
        is_male=True,
        backbone=backbone,
        pooler=pooler,
        gender_embed=gender_embed,
        classifier=classifier,
        config=config,
        device=device,
    )
    print(f"Predicted bone age (TTA): {age_tta:.1f} months ({age_tta/12:.1f} years)")

Training Details

Data

Dataset: RSNA Pediatric Bone Age (14,236 images)
Split: 12,611 train / 1,625 validation (stratified by age bins, no leakage)
Labels: Bone age in months (0-228), converted to Gaussian distributions for DLDL

Hyperparameters

Parameter	Value
Base model	google/medsiglip-448
Adaptation	DoRA (use_dora=True)
LoRA rank	16
LoRA alpha	32
LoRA dropout	0.05
LoRA targets	q_proj, k_proj, v_proj, out_proj, fc1, fc2
DLDL sigma	12.0 months
Label smoothing	0.01
Optimizer	AdamW (fused)
Learning rate	5e-5
LR schedule	Cosine with warmup
Warmup ratio	0.05
Weight decay	0.01
Batch size	8 × 2 gradient accumulation = 16 effective
Epochs	20
Precision	bfloat16
Gradient checkpointing	Enabled

Augmentation

Rotation: ±20° with constant border padding
Random brightness/contrast: ±10%, probability 0.3
Letterbox resize to 448×448 (aspect ratio preserved)

Hardware

GPU: NVIDIA RTX 4090 (24 GB VRAM)
Training time: ~2 hours
Peak VRAM usage: ~18 GB

Loss Function

KL divergence between the predicted softmax distribution and a Gaussian target:

Target: N(age_true, σ²) discretized over 228 bins, label-smoothed
Loss: KL(predicted || target) = Σ target[k] × log(target[k] / predicted[k])

Saved Artifacts

File	Description	Size
`config.json`	Model configuration (image size, hidden dim, num bins)	<1 KB
`heads.pth`	Trained pooler, gender_embed, and classifier weights	~2 MB
`adapter/adapter_config.json`	LoRA/DoRA adapter configuration	<1 KB
`adapter/adapter_model.safetensors`	LoRA adapter weights	~5 MB

The base MedSigLIP-448 model (~1.2 GB) is downloaded automatically from google/medsiglip-448 on first use.

Limitations

Trained primarily on the RSNA dataset, which lacks explicit racial/ethnic metadata. Cross-population generalization should be validated before clinical use.
The model assumes a standard left hand PA radiograph. Performance may degrade on non-standard views, right hand images, or heavily rotated inputs.
This is a research tool, not a medical device. All predictions require interpretation by qualified clinical professionals.

Ethical Considerations

Bone age assessment has significant implications in pediatric endocrinology (growth disorder diagnosis), forensic anthropology, and legal age determination for asylum cases. Automated systems must be transparent about their accuracy limitations and should never be used as the sole basis for consequential decisions.

Citation

bibtex

@misc{chronosmsk2025,
    title={Chronos-MSK: Bias-Aware Skeletal Maturity Assessment at the Edge},
    author={Rohit Rajesh},
    year={2025},
    note={Google HAI-DEF Competition Submission}
}

04RR/chronos-msk-regressor

Chronos-MSK Bone Age Regressor