Patch Ranking Map: Explaining Relations among Top-Ranked Patches, Top-Ranked Features and Decisions of Convolutional Neural Networks for Image Classification

Abstract

Since a conventional Convolutional Neural Network (CNN) using a large number of extracted features is not fully explainable and not very memory-efficient, we develop an explainable and efficient CNN model consisting of convolutional layers, a new feature selection (FS) layer, a classifier, and a novel ``Patch Ranking Map" (PRM). The PRM contains top-ranked image patches that have important associations with decisions of the CNN. Top-ranked common features selected by different FS methods are used to generate two newly defined matrices: the ``feature accumulation matrix" and the ``feature ranking matrix". Different from a typical heatmap, these two matrices are used to rank image patches in the PRM to effectively explain the relationship among an input image, top-ranked features, top-ranked feature maps, and the final classification decision. Simulation results using the Alzheimer's MRI preprocessed dataset for 4-class image classification with $6,400$ $128\times128$ images indicate that the three PRMs based on three robust top-ranked common feature sets generated by seven different FS methods have the same top two most important patches associated with Alzheimer's disease diagnosis. In addition, $8\times8$ patches of a $128\times128$ image at the 7th and 12th patch rows and at the 8th and 9th patch columns are most informative because they have the top two most important patches and the top two most top-ranked common row-wise and column-wise features. The relationship among brain regions associated with Alzheimer's disease, the top-ranked patches, the top patch rows, and the top patch columns will be analyzed based on research results in brain informatics and medical informatics. The simulations also show that the trained CNN with FS can have higher classification accuracy and smaller model size than the conventional CNN without FS. More effective and efficient optimization algorithms will be developed to select the top (most informative) features and rank patches for building an accurate and efficient CNN model with more explainable decisions that can be captured by the PRM for various image classification applications.