Weights analysis for CNN models in MRI classification
Bijen Khagi, Goo‐Rak Kwon · Alzheimer s & Dementia · 2022
Abstract Background In deep neural network learnable parameters like weights and bias of convolution kernel, normalization layer or fully connected layers (FCLs) are the one responsible for giving prediction score. Though one parameter or parameters in a layer has not decisive role, however when whole network prediction is to be done, each one has role to generate prediction value. Hence, they are important for the network performance. Method We modeled a CNN for classification and trained it using 50% of total MRI scans for 3 classes. After training the network for 50 epochs, the network reaches convergences with 100% training accuracy and around 70% validation accuracy. Then, the trained models’ parameters i.e., weights and bias of FCL layers were analyzed class wise for any correlation with its parent class. Result The weights of final FCL are plotted as in Figure 1. The correlation matrix for weights on sample MRIs with trained weights are shown in Tables 1 and 2. Here, two activation functions were used for final test results were ReLU and Leaky ReLU with final test accuracy around 67.3% and 70.9% respectively. Conclusion We attempted to study the weights pattern in FCL layer for a trained CNN model along with its correlation value for each class. [Acknowledgement] This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF‐2021R1I1A3050703). And this research was supported by the BrainKorea21Four Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (4299990114316).