Lightweight Pneumonia Detection Using Classical Descriptors on Edge AI Devices

Abstract

Pneumonia remains a pressing global health issue, and timely chest X-ray interpretation is central to effective patient care. This paper describes a lightweight pipeline for detecting pneumonia in chest radiographs using classical machine learning classifiers paired with handcrafted image features, aimed at hardware with tight processing budgets. Three widely used descriptors — HOG, SIFT, and ORB — were applied to the publicly available Kaggle chest X-ray dataset containing 5,863 labeled pediatric radiographs (4,273 Pneumonia, 1,590 Normal), pre-split into training (5,216), validation (16), and test (624) partitions. Four classifiers — Naive Bayes, SVM, Random Forest, and Logistic Regression — were trained with class-weight balancing to account for the roughly 2.7:1 class imbalance. Training and initial evaluation were carried out in Google Colab GPU-accelerated environment on a fixed held-out test set (approximately 89:11 train-to-test ratio), with 5-fold cross-validation on training data used to assess generalization. The strongest pipeline was subsequently ported to an NVIDIA Jetson Orin Nano Super for edge inference profiling. Random Forest with HOG features achieved the highest accuracy at 93.52% and an F1 score of 95.60%, while Logistic Regression with ORB features achieved the fastest inference speed at 0.187 ms per image (5,342.79 FPS). McNemar's test confirmed statistically significant gaps between the best and worst configurations (p < 0.01). Overall, the findings indicate that handcrafted features and traditional classifiers can produce clinically relevant performance within the power and memory envelopes of affordable edge hardware — a promising direction for pneumonia screening where specialist radiology support is scarce.