Detect Ankylosing Spondylitis using Deep Learning Framework

Draw bounding box on sacroiliac joint and classify ankylosing spondylitis.

If you want to know more about how to use the tensorflow object detection API, check out API_tutorial(EN) or API_tutorial_KOR(KOR)

Introduction

• Ankylosing Spondylitis(AS)

• Sacroiliac joint(SIJ)

Dataset

• 946 X-ray images(normal: 468, positive: 478)
• train: 756, test: 190 stratified(proportions of two classes were maintained), random_state=42.

Method

• Preprocessing
  : Resized to 1024X1024 maintaining aspect ratio.

• Model
  : EfficientDet-D4 1024X1024, pre-trained on COCO 2017 dataset (model_link)

• Hyperparameter (for detailed info, check out pipeline.config)
  -batch size: 4
  -optimizer: Adam
  -learning_rate_base: 0.003999999821186066
  -warmup_learning_rate: 0.000010000000474974513
  -warmup_steps: 2500
  -total_steps: 70000

• Augmentation
  : random horizontal flip, random scale cropping, random brightness adjustment, random contrast adjustment.

Results

Training Loss(Tensorboard, from step 15000 to 20000)

Classification Report & Confusion Matrix

COCO MAP score

Results

• Negative

• Positive

• Half

Reference