Medical image augmentation for lesion detection using a texture-constrained multichannel progressive GAN.

Area of Science:

• Medical imaging
• Artificial intelligence
• Computer vision

Background:

• Deep learning lesion detectors aid disease diagnosis but suffer from limited training data.
• Existing medical image augmentation methods produce low-quality synthetic images, especially for delicate textures, hindering detector performance.

Purpose of the Study:

• To propose a novel medical image augmentation method, the texture-constrained multichannel progressive generative adversarial network (TMP-GAN).
• To enhance the fidelity and utility of synthetic medical images for training deep learning-based lesion detectors.

Main Methods:

• TMP-GAN utilizes joint training of multiple channels to overcome limitations of current generation techniques.
• An adversarial learning-based texture discrimination loss is employed to improve synthesized image quality.
• A progressive generation mechanism is incorporated to enhance the accuracy of the image synthesizer.

Main Results:

• Experiments on CBIS-DDMS and a pancreatic tumor dataset demonstrated significant improvements in lesion detector performance.
• Precision, recall, and F1-score showed notable increases when detectors were trained on TMP-GAN augmented data compared to other methods.
• The Free-response Receiver Operating Characteristic (FROC) curve analysis confirmed superior performance over contrast-augmented datasets.

Conclusions:

• TMP-GAN is a practical and effective technique for medical image augmentation.
• The method significantly improves the performance of deep learning-based lesion detectors by addressing data scarcity and enhancing image quality.
• TMP-GAN facilitates efficient lesion detection case studies and advancements in medical diagnostics.