Tuning deep learning hyperparameters for magnetic resonance fingerprinting recognition

We present a novel deep learning approach for breast cancer detection. In the proposed method we simulate the magnetic resonance fingerprints obtained from pseudorandom radiofrequency pulse sequences using the population average breast tissue composition. We analyze the magnetic resonance response using deep neural networks. Two types of network architectures are considered: fully connected networks and residual-type networks. A number of networks hyperparameters sets are considered and network performance is assessed. A maximum accuracy of 0.82 and 0.81 with fully connected and residual networks is achieved with under the selected hypermarameter sets, which exceeds the accuracy with suboptimal training conditions by more than 0.30. After the hyperparameters were set, the classification accuracy exceeded 0.95.