MobileConvNeXt: An Application for Aortic Dissection Detection Using Computed Tomography Images

Abstract

This study introduces MobileConvNeXt, a compact mobile CNN for biomedical image classification, and validates it on a newly created CT aortic dissection dataset. The network is a mobile-tailored ConvNeXt variant that combines Batch Normalization (stable optimisation), Swish (smooth gradient flow), and Squeeze-and-Excitation (channel attention) to improve robustness under low compute. To further boost performance, we design a deep feature engineering (DFE) layer on top of the same backbone: two feature sets are extracted via GAP + dropout, reduced using four feature selectors plus their intersections, classified with SVM and kNN, and fused using Iterative Majority Voting with greedy search. Across 78 total outcomes, the best configuration is selected as the recommended model, where MobileConvNeXt and its DFE variant achieve 92.44% and 97.54% test accuracy, respectively, with only similar to 4.3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sim }4.3$$\end{document} M parameters, demonstrating high accuracy at low computational cost.