Optimized Training Pipeline for Deep Learning Applications in Medical Image Processing

Deep learning has revolutionized the field of digital image processing. However, training a Convolutional Neural Network (CNNs) requires a complex pipeline consisting of image normalization, data augmentation, sample mining, parameter updates, performance evaluation and monitoring. Regardless of the image processing task, development of new approaches requires this pipeline to work before any experiments can be performed. For tomographic image data, special care is necessary with regard to the modality-specific image properties. The work presented in this thesis provides a training pipeline based on the commonly used TensorFlow library. The pipeline is tailored to three medical image processing tasks: image regression, semantic segmentation and image classification. It was utilized to train CNNs in four studies on medical deep learning. In an initial study the pipeline was used to train CNNs for limited angle artifact correction in circular tomosynthesis. The CNNs were trained on simulated data and were subsequently able to correct artifacts in synthetic and real scans. On the real data an artifact reduction of 30 to 40% was achieved using a 3D ResNet. In a second study intra-individual volume change analysis in serial T1-weighted magnetic resonance imaging scans of the brain was realized with a 3D U-Net. The results demonstrated that the deep learning version could approximate the complex Voxel-guided Morphometry mapping at high quality (structural similarity index measure = 0.9521±0.0236) while reducing the computation time by 99.62%. In a third study, the pipeline was applied to vessel segmentation in contrast enhanced computed tomography (CT). Ratio-based sampling was proposed to counter the class-ratio imbalance. Using the pipeline, 2D and 3D versions of the U-Net, the V-Net and the DeepVesselNet were trained. Well performing networks were combined into an ensemble. The method achieved Dice similarity coefficients of 0.758±0.050 (veins) and 0.838±0.074 (arteries) on the IRCAD data set. Application to the BTCV data set showed a high transfer ability. In the final study, the pipeline was used to train several CNNs to classify whether CT images show an abdominal aortic aneurysm. Across the whole data set the algorithm achieved an accuracy of 0.856 and area under the receiver operating characteristic curve of 0.926. Using layer-wise relevance propagation, relevance maps were generated that offer interpretable visualization of the CNN's decision process. The presented framework enables fast prototyping of deep learning applications for medical image processing. Due to the modular design individual components can be switched easily. It is a valuable tool in the development of clinically relevant artificial intelligence algorithms