Title: A Lightweight and Interpretable Three-Hidden-Layer Backpropagation Neural Network for Mushroom Toxicity Classification Using Morphological and Olfactory Features
Authors: Nesreen Samer El_Jerjawi, Samy S. Abu-Naser
Volume: 10
Issue: 7
Pages: 105-116
Publication Date: 2026/07/28
Abstract:
Accurate identification of poisonous mushroom species is a long-standing food-safety challenge, since many toxic and edible species share highly similar morphological characteristics that are difficult to distinguish through casual visual inspection. This study proposes a lightweight and fully interpretable artificial neural network (ANN), developed entirely within the JustNN environment, for the automated binary classification of mushroom specimens as edible or poisonous using 21 standardized morphological and olfactory features. The proposed model employs a compact 21?2?1?2?1 feedforward architecture trained using the Backpropagation algorithm with momentum (learning rate = 0.6, momentum = 0.8) over 6,950 learning cycles on 2,000 training examples, and evaluated on an independent validation set of 1,000 specimens drawn from a perfectly balanced dataset of 3,000 records (1,500 edible, 1,500 poisonous). The model achieved a validation accuracy of 97.80%, correctly classifying 978 of 1,000 validation specimens, with an average training error of 0.027222 against a target error of 0.010000. To complement this result with class-level diagnostics, an architecture- and hyperparameter-matched independent replication was conducted, yielding Accuracy = 97.00%, Sensitivity = 96.80%, Specificity = 97.20%, Precision = 97.19%, F1-Score = 96.99%, and AUC-ROC = 0.9721 on the same held-out split, with five-fold stratified cross-validation confirming stability (Accuracy = 97.17% ± 0.55%, AUC-ROC = 0.9537 ± 0.0103). JustNN's native attribute-importance ranking identified gill_spacing, gill_size, and spore_print_color as the three most influential morphological features, a finding consistent with established mycological knowledge regarding the taxonomic significance of gill structure and spore characteristics in species identification. These results demonstrate that a minimal, highly interpretable neural architecture can achieve near-ceiling classification performance on a well-separated morphological dataset, offering a practical and transparent decision-support tool for food safety and mycological education applications.