Room: Exhibit Hall | Forum 1
Purpose: To improve the performance of neural network for removing various types of CT ring artifacts effectively in post-processing domain.
Methods: First section: We detected the image edges by specific morphological edge detection operator based on ring artifact images. Next we did enhancement processing by gray level transformation. The artifacts were then located by morphological hole filling and pixel connectivity method. A hybrid of bicubic and polynomial fitting interpolation was performed to compensate artifact region for different types of artifacts. Second section: We designed enhanced Fuzzy-Radial Basis Function Neural Network (RBFNN) system. Here, we selected 2 pixels which are neighbor to a certain interpolated data as inputs of our system and used interpolated data as training sample, then we trained the system by Gravitational search algorithm (GSA). Based on this, we used 15 sets of input data (all are from the processing data) to get 15 output data, then the final output data were obtained by weighing normalization processing. At last, we repaired the whole interpolated data. We also trained the system by classic error back propagation algorithm (EBPA).
Results: We compared interpolation method, interpolation-EBPA and interpolation-GSA by subjective observation (4 groups of clinical data) and objective analysis (mean gradient (MG) and SNR). The interpolation method generated worst results, where the eliminations of artifacts were not realistic by showing lowest SNR and MG values. Comparing with interpolation-EBPA, proposed method provided better visual effect (no new artifacts, high image resolution etc.) and higher metrics.
Conclusion: Through our system, the interpolation algorithm is improved and updated: complicated ring artifacts are removed completely. With application of GSA to the training of neural network, the performance of GSA is superior to that of EBPA.