Deep Convolutional Neural Networks for Computer-Aided Detection: CNN Architectures, Dataset Characteristics and Transfer Learning

Motivated by the scarcity of comprehensively annotated medical imaging datasets, the authors systematically study three previously understudied factors in applying deep convolutional neural networks to computer-aided detection. They compare CNN architectures spanning roughly five thousand to 160 million parameters and varying depths, evaluate how dataset scale and spatial image context affect performance, and analyze when and why fine-tuning models pre-trained on ImageNet transfers usefully to medical tasks.

Using two concrete problems, thoraco-abdominal lymph node detection and interstitial lung disease classification, the study achieves state-of-the-art performance on mediastinal lymph node detection and reports the first five-fold cross-validation classification results for predicting ILD categories from axial CT slices. The empirical analysis and design insights were positioned as guidance transferable to building high-performance CAD systems for other medical imaging tasks.