Dipartimento di Ingegneria informatica, automatica e gestionale

Pneumonia is one of the most relevant and dangerous chest diseases in the world. It can be detected through images generated by Chest X-Rays (CXR) or CT scans. In recent years, several computer-aided diagnosis (CAD) systems have been developed, especially aided by recent advances in deep learning. One of the most recent innovations is the Kolmogorov-Arnold Network (KAN) which was proposed as an alternative to the classical Multi-Layer Perceptron (MLP): KANs outperform MLPs in terms of accuracy and interpretability. The concept of learning activation functions was applied also in convolutional layer: these are called Convolutional KANs (CKANs). At the moment, deep learning-based CAD systems use convolutional layers to extract features from images and the MLP classifier for classification tasks, while in this study an innovative solution using Convolutional KAN layers and KAN classifier is proposed. The aim of this study is to show how the most recent novelties in deep learning, KANs and CKANs, are better alternatives than the classical solutions in terms of performance. For this reason a deep neural network called 'JAM-net' is presented for CXR binary classification tasks (normal/pneumonia). JAM-net is derived from the actual state-of-the-art model for pneumonia detection, a Fuzzy Attention-aided Deep Neural Network called 'FA-net', enriched by the CKAN and KAN layers. The new JAM-net model is able to overcome the actual state-of-the-art performance with 98.84% of accuracy.