A CNN model with pseudo dense layers: some case studies on medical image classification

Evolution of deep Convolutional Neural Networks (CNNs) has primarily focused on enhancing the convolutional layers of various deep neural network models. Throughout this evolution, the fully connected multi-layer neural networks (also known as dense layers) in these models have largely remained unchanged. However, dense layers overlook the potential for feature planes generated from convolutional layers to selectively influence the class memberships of input images. This paper introduces a novel alternative termed P-Net, or pseudo dense layers, to the conventional fully connected dense layer found in various deep CNN models. P-Net leverages the spatial concentration of features within each feature plane for different image classes. Furthermore, P-Net incorporates weight sharing within feature planes, which introduces an inductive bias and reduces the parametric complexity of the model. We evaluated our method on two publicly accessible benchmark medical image datasets: the Lung and Colon Cancer Histopathological dataset (LC25000) and the Brain Tumor MRI dataset. To rigorously assess P-Net, we compared it against three widely used standard deep CNN models: ResNet, DenseNet, and EfficientNet. P-Net achieved the highest accuracies of 99.2%, 100%, and 100% on the Brain Tumor MRI, Lung, and Colon datasets, respectively. Across all cases, deep CNN models incorporating P-Net consistently outperformed conventional deep CNN models on these datasets, demonstrating the effectiveness of our approach in medical image classification tasks. The source code of this work is available at: https://github.com/mridulbiswas402/P-Net.git.