A Multi-Cluster Random Forests-Based Approach to Super-Resolution of Abdominal CT Images Using Deep Neural Networks
Mahdieh Akbari, Amir Hossein Foruzan, Yen‐Wei Chen · 2020
To enhance the resolution of abdominal computed tomography slices, a super-resolution method based on random forests is proposed. We classify input LR patches using an Auto-Encoder network, and we employ several random forest machines for the estimation of HR data. The random forest machines are trained, and they are used to reconstruct HR patches in the test phase. Compared to the conventional K-means clustering and interpolation techniques, our results improved the performance of the SR algorithm by at least 0.02 and 5.7% using the SSIM and PSNR measures. While the conventional RF-based method generated unwanted spots in the HR image, the quality of our results outperforms the standard RF-based approach. Moreover, the proposed intensity profile follows that of the original image and preserves the edges better; however, the conventional technique generates false high-frequency oscillations.