Mammo-GAN-Assisted Deep Network Training Scheme for Lesion Detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning models for lesion detection require extensive training data, especially borderline cases, which are difficult to acquire.
• The performance of lesion detection networks is often limited by the scarcity of challenging, borderline training examples.

Purpose of the Study:

• To develop a novel method for augmenting training datasets by simulating and removing lesions to create more challenging cases for deep learning models.
• To enhance the performance of lesion detection deep networks by increasing the number of borderline cases through data augmentation.

Main Methods:

• A Cycle-GAN-based Lesion Simulator (LS) and Lesion Remover (LR) were developed using a mammography dataset.
• LS generates lesions in normal patches, and LR removes lesions from abnormal patches, creating harder-to-distinguish cases.
• The LS-LR model was trained at different epochs (25%, 50%, 75%) to control the simulation impact, and a ResNet18 model was retrained using the augmented data.

Main Results:

• Retraining ResNet18 with data augmented by LS-LR at 50%/75% training epochs significantly improved lesion detection performance (AUC=0.901) compared to the baseline (AUC=0.870) on a mammography test set.
• External validation demonstrated generalizability, with AUC improvements on an independent mammogram dataset (0.866 vs. 0.839) and a chest X-ray dataset (0.975 vs. 0.964).

Conclusions:

• The proposed LS-LR model effectively transforms existing medical imaging data into borderline cases, thereby improving the robustness and performance of deep learning-based lesion detection.
• This data augmentation strategy offers a promising approach to enhance diagnostic accuracy in mammography and chest X-ray interpretation, addressing the challenge of limited borderline case availability.