Near-pair patch generative adversarial network for data augmentation of focal pathology object detection models.

Area of Science:

• Medical imaging
• Machine learning in diagnostics
• Artificial intelligence in healthcare

Background:

• Limited medical training data hinders machine learning for diagnostic applications.
• Object detectors require extensive labeled images, which are costly and time-consuming to acquire.
• Generating synthetic pathology-present images can augment datasets for training.

Purpose of the Study:

• To develop a method for distant supervision of object detectors using synthetic pathology-present labeled images.
• To reduce the challenges associated with curating large volumes of medical training data.
• To improve the performance of object detectors in medical diagnostic tasks.

Main Methods:

• Utilized a cyclic generative adversarial network (cycleGAN) with innovations in training data selection (near-pairs) and realism metric integration (Fréchet inception distance).
• Trained the model on pediatric chest radiograph patches, generating synthetic fracture-present images.
• Augmented a real dataset with synthetic images to train an object detector (YOLOv5).

Main Results:

• An observer study showed synthetic fracture-present images were rated as more likely to be real fractures compared to fracture-absent images.
• Object detector (YOLOv5) trained on a mix of real and synthetic data achieved higher recall (0.57) and F2 score (0.59) than one trained on real data alone (recall 0.49, F2 score 0.53).

Conclusions:

• The proposed method successfully generates visually realistic pathology.
• The synthetic data significantly improved object detector performance for rib fracture detection.