One‑Class Deep Learning Framework for Anti‑NTD Plant Identification Using Synthetic Negatives and Transfer Learning

The discovery of plant‑derived therapies for Neglected Tropical Diseases (NTDs) is hindered by the scarcity of authentic negative samples for training robust classification models. This study presents a one‑class deep learning framework that overcomes this limitation by generating a synthetic negative class, thereby enabling accurate binary classification without real non‑active specimens. The EfficientNetB4 architecture, which was pre‑trained on ImageNet, was fine‑tuned using a two‑phase protocol on a dataset of 4,950 images from 99 anti‑NTD plant species augmented with 990 synthetically generated negatives. A focal loss function (γ=2.0, α=0.25) addressed the inherent 5:1 class imbalance, and botanical‑specific augmentations (rotation, flipping, brightness/contrast adjustment) enhanced model generalisation. The proposed system achieves 94.4% classification accuracy (95% CI: 93.2–95.4%) with an AUC‑ROC of 0.981 and an F1‑score of 0.955, demonstrating excellent discriminative performance. Confidence scores correlate strongly with prediction accuracy (Pearson’s r=0.891, p<0.001), enabling practical uncertainty quantification. The model processes images at 25.8 frames per second on a Tesla T4 GPU, making it suitable for high‑throughput screening. Error analysis revealed that 46.6% of misclassifications occur at confidence levels below 0.6, with pressed/dried specimens (42%) and complex natural backgrounds (38%) as primary error sources. This work establishes that morphological features from standard plant photographs can be effectively leveraged for therapeutic property prediction, offering an accessible, open‑source computational tool to accelerate medicinal plant prioritisation, reduce dependency on scarce expert knowledge, and support conservation efforts for promising species. The complete code and trained model are publicly available to facilitate adoption and further validation by the botanical and pharmacological communities.