|
One of the significant part of CNN is feature extraction module. For computer vision, image’s patterns are extracted by filters and convolution operations in feature extraction. The CNN learns filter weights from signal data and extracted features. So, widely used filters are based on Gaussian distribution initializing, together with the low-level feature extraction in state-of-the-art architectures. According to CNN, reducing parameters and computing resources while accuracy maintaining is challenging. Most CNN compression research aims at general image which inappropriate for medical image. The MRNet dataset with MRNet CNN model for knee MRI diagnosis assisting by Bien et al. from stanfordML group, have recently been developed. This MRNet achieved AUC of 0.965 on internal ACL tear dataset classification. Based on MRNet, we modified feature extraction module to compress MRNet with measured on ACL tear within MRNet dataset. We designed with a foundation of 2 n form filter, and supplemented by MRI-cut selection. We split MRI by cuts and tested results of each cut combination together. The combination leading to the best accuracy is Coronal/ Sagittal, we used it as input dataset. Then we replaced the filters in baseline model by 2×2 filters, 4×4 filters, medical DIP 8×8 filters, with symmetric padding to eliminate shift problem in even-sized filter. The MRNet baseline (trained from scratch) got average error rate as 8.50% and our proposed got 12.94% but a number of parameters is pruned by 52.250%, a number of FLOPs is pruned by 46.145%, and requiring only Coronal and Sagittal.
|
