Deep fused explainable neural architecture for the automatic recognition of therapeutic medicinal plants
Clinical Snapshot
PICO Framework
| P — Population | Medicinal plant leaf images across multiple botanical species (30 classes in primary dataset; additional Bangladeshi and MedLeaves datasets used for external validation) |
| I — Intervention | MedPlantNet — a deep fused explainable neural architecture combining fine-tuned Deep Convolutional Neural Network (DCNN) feature extraction with a customised Support Vector Machine (SVM) classifier, augmented by Grad-CAM and Occlusion Sensitivity visualisation |
| C — Comparator | Existing state-of-the-art (SOTA) deep learning and machine learning algorithms for medicinal plant image classification (specific comparator models not enumerated in abstract) |
| O — Outcomes | Classification accuracy, specificity, sensitivity, and Area Under the Curve (AUC) for medicinal plant leaf identification; model interpretability via explainability visualisation |
Bottom Line
MedPlantNet is a proof-of-concept deep learning system for automated medicinal plant leaf classification, combining convolutional neural network feature extraction with a support vector machine classifier and explainability visualisation. While the hybrid architecture and multi-dataset evaluation represent methodological effort, the reported performance figures — 99.87% accuracy and an AUC of 1.0 — are implausibly high and raise serious concerns about overfitting, data leakage, or benchmark saturation. No confidence intervals are provided, no real-world or prospective validation is performed, and the clinical translation pathway is entirely absent. The study does not address misclassification consequences, which are potentially serious in a domain where toxic plant misidentification carries direct patient harm risk. For Australian clinicians and pharmacists, the model has no demonstrated relevance: it is not trained on Australian flora, is not aligned with TGA quality assurance frameworks, and has not been evaluated under field conditions. This work represents an early-stage engineering contribution that requires substantial independent validation, prospective real-world testing, and regulatory engagement before any clinical or quality assurance application could be considered. Senior clinicians should not alter practice based on these findings.
Key Findings
P Value: Not reported
Effect Size: Validation accuracy of 99.87% on the Mendeley Medicinal Plant Leaf Database; 99.27% on the Bangladeshi medicinal plant dataset; 99.80% on the MedLeaves dataset
Primary Outcome: Automated classification of medicinal plant leaf images across 30 species classes using the MedPlantNet architecture
Nnt Or Sensitivity: Sensitivity reported as a metric but specific values not provided in the abstract; AUC reported as 1.0 (perfect discrimination), which is clinically implausible and methodologically suspect
Confidence Interval: Not reported — a critical methodological omission
Clinical Application
Computational feasibility is demonstrated on benchmark datasets. Real-world feasibility requires: (1) deployment-ready software with validated image capture protocols, (2) integration with pharmacy or clinical information systems, (3) regulatory approval for any diagnostic or quality assurance application, and (4) prospective field validation with non-expert image capture. None of these are addressed. Limited direct applicability to Australian practice. The TGA regulates listed and registered complementary medicines under the Therapeutic Goods Act 1989, and herbal ingredient identification is subject to pharmacopoeial standards (e.g., British Pharmacopoeia, Australian Pharmacopoeia). The model is not trained on Australian native medicinal species or TGA-relevant botanical taxa. The RACGP does not currently endorse automated plant identification tools for clinical use. PBS listing of herbal medicines is minimal, and quality assurance frameworks for complementary medicines in Australia (TGA Listed Medicines pathway) would require validated, regulatory-grade identification methods — a bar this study does not approach. Indigenous Australian plant medicine contexts would require entirely separate ethnobotanical and cultural considerations. Potentially applicable to botanists, pharmacognosists, herbal medicine quality assurance personnel, and pharmacy compounding contexts where rapid plant species identification is required. Not validated for direct patient care decisions.
Abstract
BACKGROUND: Medicinal plants have been utilized for centuries for their beneficial properties and significant impact on traditional medical systems worldwide. Indian medicinal plants and their therapeutic properties have recently gained increased attention from the global healthcare community. Accurate identification and classification of medicinal plants ensure proper therapeutic applications. However, botanists manually identify these plants, which is subjective, labor-intensive, and time-consuming. OBJECTIVE: The main objective of this study is to develop an accurate and automated classification of these medicinal plants for diverse applications such as conservation initiatives, quality assurance, and the formulation of novel herbal treatments. METHODS: A Deep Fused Explainable Neural Architecture termed MedPlantNet is proposed for robust recognition of medicinal plant leaves. The deep discriminative features are extracted from the medicinal plant images using a fine-tuned Deep Convolutional Neural Network (DCNN). A customized Support Vector Machine is developed to efficiently classify these extracted deep features. To enhance model interpretability, a meta-visualization framework incorporating Grad-CAM and Occlusion Sensitivity visualization is implemented to identify the regions of the image that contribute to the classification decisions. We utilized the Mendeley benchmark Medicinal plant leaf database with 30 classes to train and evaluate the model. RESULTS: To assess the effectiveness of the proposed model, accuracy, specificity, sensitivity, and AUC are calculated. The proposed system outperforms existing state-of-the-art (SOTA) algorithms for medicinal plant classification, achieving an optimal maximum validation accuracy of 99.87% and AUC of 1. To further validate the generalizability external validation was performed on Bangladeshi medicinal plant dataset and MedLeaves dataset where the proposed model MedPlantNet attained 99.27% and 99.80% validation accuracy respectively. CONCLUSION: MedPlantNet provides an automatic framework for medicinal plant identification based on deep feature extraction, classification and explainable visualization techniques. The results obtained indicate its potential utility in supporting biodiversity conservation, quality assurance and herbal medicine research.
References
- 1.Govindaraj, P., & Sasikaladevi, N. (2026). Deep fused explainable neural architecture for the automatic recognition of therapeutic medicinal plants. Daru: Journal of Faculty of Pharmacy, Tehran University of Medical Sciences. https://doi.org/10.1007/s10462-023-10631-z
This content is for educational purposes for healthcare professionals only and does not constitute clinical advice. Clinical decisions should be based on individual patient assessment, current guidelines, and appropriate specialist consultation. Editorial Standards · Privacy Policy · Terms of Service