Modelpublic

ASJ234/retinanet

sourceHugging Faceupdated 2mo agoView on Hugging Face

0likes6downloads

Model Card

RetinaNet-ResNet50-FPN-V2 — TB Lesion Detection

Model Description

This model is a RetinaNet with FPN for detecting Tuberculosis (TB) lesions on chest X-ray images from the TBX11K dataset.

—Architecture: RetinaNet with FPN
—Backbone: ResNet-50 + FPN v2
—Framework: torchvision
—Number of classes: 2

Classes

ID	Class
1	ActiveTuberculosis
2	ObsoletePulmonaryTuberculosis

Training Details

—Dataset: TBX11K (chest X-rays)
—Epochs: 75
—Batch size: 4
—Optimizer: AdamW
—Learning rate: 0.0001
—Weight decay: 0.0001
—Warmup epochs: 3
—EMA decay: 0.99
—Gradient clipping: 10.0

Augmentation

—Horizontal flip: 0.5
—Brightness: +/-0.3
—Contrast: +/-0.3
—Gamma: 0.2
—Noise std: 0.05

Performance

Best mAP@0.5:0.95: 0.0658 (epoch 75)

Metric	Value
mAP@0.5:0.95	0.0658
mAP@0.5	0.1739
mAP@0.75	0.0264
mAP_small	0.0000
mAP_medium	0.0514
mAP_large	0.0758
AR@1	0.1008
AR@10	0.2193
AR@100	0.2461

Usage

Loading Weights

python

import torch

# Load the EMA weights (recommended) or best_model weights
state_dict = torch.load('weights/ema_model.pth', weights_only=True)
model.load_state_dict(state_dict)

Files

weights/
  ema_model.pth        # EMA weights (recommended for inference)
  best_model.pth       # Best weights by validation mAP
  last_checkpoint.pth  # Full checkpoint (includes optimizer state)
config.json           # Training configuration
metrics.json          # Validation metrics (COCO evaluation)
metrics_tta.json      # Test-time augmentation metrics
confusion_matrix.png  # Confusion matrix visualization
curves/               # Training curves
explain/              # Grad-CAM / attention visualizations

Citation

bibtex

@misc{tbx11k_detection,
  title={TB Lesion Detection on Chest X-rays},
  year={2024},
  note={TBX11K Dataset},
}