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