Mammo-GAN-Assisted Deep Network Training Scheme for Lesion Detection.
Authors
Affiliations (3)
Affiliations (3)
- Department of Radiology, The University of Pittsburgh, 200 Lothrop Street, Pittsburgh, PA, 15237, USA. [email protected].
- Department of Industrial Engineering, The University of Pittsburgh, 1025 Benedum Hall, Pittsburgh, PA, 15261, USA. [email protected].
- Department of Radiology, The University of Pittsburgh, 200 Lothrop Street, Pittsburgh, PA, 15237, USA.
Abstract
The performance of lesion detection deep networks is determined by how well the network can learn key features from borderline cases, but finding such cases is challenging. We propose to increase the number of borderline cases for a given network by converting existing training data through applying a Cycle-GAN-based Lesion Simulator (LS) and Lesion Remover (LR). We developed our LS-LR using the developmental portion (80%) of a 10,414 mammography patch dataset from 4789 unique patients (2416 with lesions and 2312 normals). LS creates lesions in normal patches and LR removes lesions from lesion patches, producing harder-to-distinguish cases. Importantly, training LS-LR at lower training epochs (25%, 50%, 75%) can modify the simulation impact. We trained a ResNet18 using a training set for a lesion detection task (baseline). Using our LS-LR at different training epochs, we converted the training samples to be more challenging and retrained the ResNet18 from scratch using the updated set. Our LS-LR at 50%/75% training epochs created the most effective samples for retraining ResNet18, and it achieved superior performance (AUC = 0.901) compared with the baseline (AUC = 0.870; p = 0.00005) on the test mammography dataset. External validation confirmed generalizability: for an independent mammogram dataset (134 lesions, 180 normals), AUC improved to 0.866 from 0.839 (p = 0.007); for a chest X-ray dataset (826 nodules/masses, 1642 normals), AUC improved to 0.975 from 0.964 (p = 0.011). The results showed that our LS-LR model improved lesion detection performance in mammograms and chest X-ray images by transforming existing data into borderline cases.