Few-shot learning for classification of SEM images from green-synthesized nanoparticles of Momordica cymbalaria

Abstract

Nanotechnology offers wide range of applications owing to the unique physicochemical and biological properties of nanoparticles. Synthesis of nanoparticles from plant extracts provides a sustainable and safe alternative to the conventional chemical methods. In the current research phyto-synthesis and characterization of silver nanoparticles (AgNPs) and calcium carbonate nanoparticles (CaCO₃ NPs) using fruit and root extracts of Momordica (M.) cymbalaria, is reported. The characterization was carried out using Ultraviolet (UV)–Visible spectroscopy, Fourier Transform Infrared (FTIR) spectroscopy, Scanning Electron Microscopy (SEM), and Energy-Dispersive X-ray (EDAX) spectroscopy. EDAX confirmed CaCO₃ NPs to be rich in oxygen (42–45%) and calcium (30%), while AgNPs exhibited distinct silver peaks (15–20%) along with significant carbon and oxygen content. UV–Visible spectroscopy and FTIR spectra confirmed functional groups and surface plasmon resonance peaks characteristic of the synthesized nanoparticles. The SEM analysis revealed spherical CaCO₃ nanoparticles (200–600 nm), whereas AgNPs displayed irregular aggregates (3–25 μm). To address the limited dataset of SEM images, a few-shot learning (FSL) framework was adopted for nanoparticle classification across four classes (AgNO₃-root, AgNO₃-fruit, CaCO₃-root, CaCO₃-fruit). MobileNetV2 and ResNet50 were trained for classification of SEM images. MobileNetV2 with episodic training, Mahalanobis distance, and KMeans clustering (Configuration 3) achieved 95% accuracy with a compact 8.98 MB as model size, outperforming baseline transfer learning approaches. The findings highlight the potential of automated classification of nanoparticles’ SEM images using FSL under low-data regimes.

Introduction

Nanotechnology has become an integral part of modern science, with far-reaching applications in medicine, drug delivery, electronics, cosmetics, and food processing¹. A recent review comprehend on dimensions, characterization, physicochemical properties and the applications for biologist on nanoparticle². Owing to their high surface-area-to-volume ratio and nanoscale dimensions, nanoparticles display unique size-dependent physicochemical, optical, magnetic, and mechanical properties that make them attractive candidates for various biomedical and industrial applications³.

With the advances in nanotechnology, alternative methods of nanoparticle synthesis that are sustainable and environmentally safe are becoming increasingly important. Though the conventional chemical based nanoparticle synthesis is effective, it involves usage of toxic reagents and generation of hazardous by-products. This has led to exploration of greener alternatives, especially plant based synthesis of nanoparticles that offer a non-toxic, cost-effective, and scalable approach⁴. Plant extracts, rich in bioactive compounds such as flavonoids, terpenoids, polysaccharides, proteins, and alkaloids, act as natural reducing and stabilizing agents, facilitating the synthesis of nanoparticles with well-defined morphology and enhanced biocompatibility^5,6.Amongst the diverse biomedical applications of biosynthesized nanoparticles, their role in the management of Type 2 Diabetes Mellitus has been crucial⁷ and plants such as Momordica genus are well known for glucose regulation and metabolism, thus improving insulin sensitivity. However, the clinical translation of these phytoconstituents is often limited by poor bioavailability and inadequate targeted delivery. In this regard, the synthesis of phyto nanoparticles offers a promising strategy to enhance therapeutic efficacy by improving stability, bioavailability, and site-specific action of these critical antidiabetic compounds.

Phytonanoparticles (PNPs) synthesized using medicinal plants have demonstrated superior biocompatibility and enhanced biological activity. For example, silver nanoparticles (AgNPs) derived from Momordica charantia and Azadirachta indica have shown significant antidiabetic effects, including improved insulin sensitivity and strengthened antioxidant defence mechanisms^8,9. Similarly, phyto-synthesized calcium carbonate (CaCO₃) nanoparticles—an emerging and relatively underexplored class—have exhibited promising potential in enhancing insulin secretion and protecting pancreatic β-cells¹⁰. Nanoparticles synthesized from plants such as Ocimum sanctum and Trigonella foenum-graecum have further been reported to reduce oxidative stress and promote pancreatic tissue regeneration^11,12. There are reports on green synthesis of AgNPs from Momordica cymbalaria and their role in antidiabetic activity¹³. But studies on calcium carbonate (CaCO₃) nanoparticles and their complete characterization from Momordica cymbalaria is limited .

In parallel with advances in plant-mediated nanomaterial synthesis, computational approaches—particularly Deep Learning (DL), a subset of Machine Learning (ML)—are emerging as powerful tools in nanomedicine for the automated analysis of complex experimental datasets¹⁴. In the context of phyto-synthesized nanoparticles, DL-based image classification of SEM images offers a robust approach for nanoparticle characterization. SEM images capture critical morphological attributes—including particle size, shape, aggregation state, and surface texture—which are key determinants of biological performance, cellular uptake, and therapeutic efficacy¹⁵. By leveraging these features, DL algorithms can effectively differentiate between nanoparticle types and synthesis conditions, thereby facilitating precise quality assessment and reproducibility of biosynthesized nanomaterials, including Momordica-derived AgNPs and CaCO₃ nanoparticles.Supervised deep learning techniques have demonstrated high accuracy in nanoparticle classification when trained on large, well-annotated datasets. However, in experimental nanomedicine—particularly for phyto-nanoparticles such as Momordica cymbalaria-mediated CaCO₃ nanoparticles—data scarcity and class imbalance remain significant challenges. This limitation has driven growing interest in few-shot learning (FSL), which aims to generalize knowledge from a limited number of labelled examples¹⁶. The current research aims at :

• i)Synthesis and characterization of silver and calcium nanoparticles using fruits and root extracts of Momordica cymbalaria.
• ii)Algorithm for N-way-K shot Episodic training of SEM image classification of Phyto-Synthesized Nanoparticles.
• iii)Evaluation of the proposed FSL method (Prototype + Mahalanobis + KMeans + Episodic Training) for SEM image classification by t-SNE plots, ablation study and paired t-test.

Related work

Deep convolutional neural networks (CNN) was employed in the study proposed by¹⁷, which combined the transfer learning techniques, specifically using VGG16 and ResNet50 architectures, for the classification of SEM images of pharmaceutical excipients. The highest classification accuracy of 96.9% was achieved using the VGG16 model with a batch size of 48 and a learning rate of 1 × 10⁻³. In comparison, the ResNet50 model attained a slightly higher accuracy of 97% when trained with a batch size of 16 and a learning rate of 1 × 10⁻². The study¹⁸ proposes SEM-Net, which combines deep features from pre-trained AlexNet and ResNet-50 with Binary Particle Swarm Optimization (BPSO) for feature selection and classification using a Support Vector Machine (SVM). Evaluated on a custom SEM image dataset of Ag-doped SnO₂ nanoparticles, the method achieved 99.3% accuracy using 3112 optimized features, highlighting its effectiveness in handling complex SEM image classification. A CNN model based on VGG-16 with transfer learning was applied by¹⁹ to extract and classify particle morphology features of 67 pharmaceutical excipients from SEM images. Using uniform manifold approximation and projection (UMAP) for dimensionality reduction and hierarchical clustering analysis (HCA), the excipients were grouped into seven clusters with similar morphologies, demonstrating the CNN’s effectiveness in identifying structural similarities. A hybrid Vision Transformer–ResNet50 (ViT-ResNet50) framework²⁰, integrated with semi-supervised learning and SHAP-based explainability, was proposed for efficient and interpretable SEM image classification. It achieved classification accuracies of 93.65% on the Aversa dataset and 84.76% on the UltraHigh Carbon Steel Database, surpassing baseline models and demonstrating effectiveness in limited labelled data scenarios. MobileSEMNet, an enhanced deep learning model built on MobileNetV2 and optimized using Neighborhood Component Analysis and an SVM classifier, was proposed for SEM image-based material classification²¹. Trained on a dataset of over 20,000 images across 10 classes, the model achieved a classification accuracy of 96.87%, outperforming the baseline MobileNetV2 model which achieved 94.85%, demonstrating its effectiveness in improving automated material recognition. While previous studies have demonstrated high accuracy in SEM image classification using CNNs, hybrid architectures, and optimization techniques, most models rely on large datasets or domain-specific configurations that may not generalize well to diverse or limited-data scenarios. Moreover, limited attention has been given to addressing class imbalance and enhancing classification robustness when working with few-shot or low-sample SEM image datasets. The research gap highlights the need for training the models for classification under low data settings without sacrificing accuracy.

Baseline of the work

1. Few-Shot Learning: FSL is an ML approach. It can be broadly classified into meta-learning and metric-based learning approaches²². Meta-learning, also known as “learning to learn,” focuses on training a model on a variety of tasks so it can quickly adapt to new, unseen tasks with very few examples. Metric-based methods learn a similarity function or a metric to compare the similarity between examples. They project data points into a feature space where similar data points are closer to each other. These methods then use this learned metric to classify new examples based on their similarity to known examples. Common examples of metric-based approaches include Siamese networks, Matching networks, and Prototypical networks. Prototypical Network, which represents each class by the mean of its embedded support samples in a latent feature space. Classification is then performed by computing the distance between the query embedding and each class prototype using a metric such as Euclidean or Mahalanobis distance. In this work, a Prototypical Network²³ based strategy has been used under the N-way-K-shot setting, by employing deep feature encoders (MobileNetV2²⁴ ResNet50²⁵) and extending the distance function to incorporate covariance through the Mahalanobis metric²⁶. This enables the representation and classification of nanoparticle SEM images, even under few-shot constraints.

2. Mahalanobis Distance: It is a covariance-aware distance metric that accounts for the scale and correlation of features in high-dimensional data, making it more suitable than Euclidean distance for comparing feature embeddings with varying distributions. Unlike Euclidean distance, which treats all dimensions equally, Mahalanobis scales differences using the inverse covariance matrix, thus normalizing feature variance and improving class separability.

3. Prototype Learning & KMeans: In prototype learning, each class is represented by a prototype—typically the mean of feature embeddings from its support samples—serving as the class centroid in the embedding space. KMeans clustering was used to derive prototypes.

4. CNN (Convolutional Neural Networks) Encoders: Pretrained CNN encoders, MobileNetV2 and ResNet50, which are trained on ImageNet²⁷ was used to encode the images. MobileNetV2 is a lightweight architecture, uses depth wise separable convolutions. ResNet50, uses skip (residual) connections. ResNet50 provides representational capacity at the expense of a bigger model size, whereas MobileNetV2 is targeted for efficiency and speed. This comparison makes it possible to compare the performance of deep and lightweight encoders in few-shot learning scenarios.

5. Leave-One-Out Cross-Validation (LOOCV): LOOCV is a strategy where, for a dataset of size N, each sample is used once as the test set while the remaining N − 1 samples form the training set, repeated for all N samples. It is particularly effective for small datasets, providing an unbiased performance estimate by maximizing training data usage in each iteration.

6. N-way-K-shot classification: In an N-way-K-shot classification setting, FSL is formulated as a meta-learning problem where the model learns to classify among N classes using only K labelled examples per class (support set). Training procedure followed for N-way-K shot classification as given in the Algorithm 1. The input to the algorithm is an image dataset D={(x_i,y_i)}_{i=1 to N}​ with labelled SEM images, CNN encoder f_θ (e.g., MobileNetV2 or ResNet50), Number of classes per episode N, Number of support examples per class K, Number of query examples per class Q, Number of training episodes T and Distance metric was Mahalanobis. Output of the algorithm is a trained encoder f_θ. For each training episode, the algorithm begins by randomly selecting N classes from the dataset D. For each class, it samples K support images (forming the support set S_c) and Q query images (forming the query set Q_c). The union of all support and query sets across classes constitutes the episodic training data. The convolutional encoder network f_θ is used to map both support and query samples into a lower-dimensional embedding space z. These embeddings serve as feature representations for downstream prototype-based classification. For each class c, the class prototype µ_c is computed using the KMeans center of support embeddings corresponding to that class. Additionally, the class covariance matrix Σ_c is estimated using the embeddings. Each query sample is compared against all class prototypes using the Mahalanobis distance d_jc, a metric that accounts for intra-class feature variance. The class with the minimum Mahalanobis distance is predicted as the query sample’s label . The negative Mahalanobis distances are passed through a softmax layer to produce class probabilities. The cross-entropy loss between predicted and true labels is computed, and the encoder parameters θ are updated via gradient-based backpropagation. After T episodes of episodic training, the algorithm outputs a trained feature extractor f_θ capable of encoding new images for few-shot classification tasks.

Algorithm 1.

N-way-K-shot episodic training framework.

Experimental methods

Green synthesis of nanoparticle

The research on Momordica cymbalaria was conducted in accordance with the IUCN Policy statement on research involving species. As the species is not categorized under threatened categories of the IUCN Red List, and only minimal, non-destructive sampling was carried out, the study does not adversely impact natural populations. For the current research, M. cymbalaria plant along with dry root and fruit was obtained from Raketla village, Uravakonda Anantapur dist. Andhra Pradesh, India (Lat: 14.8248760/Long:77.2338020). The plant sample has been previously collected and deposited in Herbarium of Government Degree College (Autonomous), Anantapur (Deposition number: GCAATP1025) in 2017. The collected plant parts (root and fruit) were washed thoroughly with distilled water, shade dried for 15 days and grounded into fine powder using mechanical grinder. The root and fruit powders (10 g) of Momordica cymbalaria were soaked in 100 mL of methanol (Analytical Grade) for component extraction. The mixture was continuously agitated at 700 rpm for five hours to ensure thorough mixing and maximize extraction efficiency. The solution was then filtered, and the resulting methanolic extract was used for nanoparticle synthesis.

To synthesize calcium carbonate nanoparticles (CaCO₃ NPs), 0.05 M CaCl₂ was added to 25mL of the prepared methanolic plant extract. The pH of the mixture was adjusted to 8.5 using NaOH and agitated for 2 h using a magnetic stirrer. The Na₂CO₃ was added, that resulted in a visible colour change a positive indicative of nanoparticle formation. The reaction mixture was sealed and incubated at 27 °C for 2–3 days to allow complete nanoparticle precipitation. The resulting precipitate was collected by centrifugation at 8000 rpm, followed by washing with distilled water and ethanol (Analytical Grade) to remove residual organic compounds and impurities. The final white precipitate was air-dried to obtain CaCO₃ nanoparticles in powdered form. For the preparation of silver nanoparticles (AgNPs), 20 ml of 0.1mM AgNO₃ solution was added to 100mL of the plant extract. The mixture was stirred for 15 min with magnetic stirrer, incubated at 27 °C in dark for 24 h. This facilitated complete precipitation and formation nanoparticles. The resulting precipitate was collected by centrifugation at 8000 rpm and washed alternately with ethanol and distilled water. These particles were air-dried for 48 h to obtain AgNPs in powdered form.

Characterization of nanoparticles

To visualize the nanoparticles’ structural features, scanning electron microscopy (SEM) was employed (ZEISS EVO 10). Prior to SEM imaging, the samples underwent a sputtering (Quorum SC7620) process to improve conductivity and image resolution. Although biogenic nanoparticles often exhibit partial conductivity due to their phytochemical capping agents, the application of a thin metallic layer remains crucial for achieving optimal imaging contrast and preventing surface charge accumulation. The SEM technique enabled high-resolution imaging of nanoparticle and their other characteristics such as size distribution, morphology, aggregation behaviour, and surface texture.

The optical characteristics and surface plasmon resonance (SPR) behaviour of the nanoparticles were assessed through UV–Visible absorption spectroscopy (GENESYS 10 S UV–VIS). A dilute ethanolic suspension (0.1%) of the nanoparticles was prepared by dispersing 0.05 mg of nano powder in 5 ml of ethanol. The spectral scans were recorded in the range of 200–800 nm. The obtained spectra provided critical insights into the optical transitions of the synthesized nanoparticles, including the characteristic absorption peaks indicative of nanoparticle formation and stability.

Energy Dispersive X-ray Analysis (EDAX), integrated with the SEM system, was utilized for qualitative and quantitative elemental profiling of the nanoparticles. This technique relies on detecting characteristic X-rays emitted from the sample surface upon interaction with a high-energy electron beam. EDAX proved particularly effective in confirming the presence of silver (Ag) and calcium (Ca) in their respective nanoparticulate forms. It further allowed for the detection of trace elements, analysis of compositional purity, and mapping of elemental distribution across the nanoparticle matrix.

To explore the surface chemistry and identify the functional groups responsible for capping and stabilizing the nanoparticles, Fourier Transform Infrared (FTIR) spectroscopy was performed (Nicolet Summit X). The samples were either directly analysed using the Attenuated Total Reflectance (ATR) mode or mixed with potassium bromide (KBr) to form transparent pellets. Background scans were initially performed to calibrate the instrument, following which the samples were scanned over an appropriate mid-infrared range (typically 4000–400 cm⁻¹). The resulting spectra were interpreted based on the position and intensity of absorption bands, allowing identification of specific functional groups such as hydroxyls, carbonyls, amines, and others associated with phytochemicals from M. cymbalaria. These functional groups are indicative of the biomolecules involved in reduction and stabilization during nanoparticle synthesis.

System architecture for SEM image classification

Figure 1, shows the system architecture of proposed method for classification of SEM images of phyto-synthesized nanoparticles. The classification dataset contains grayscale SEM images categorized into four classes—AgNP-Root, AgNP-fruit, CaCO₃-Root, and CaCO₃- fruit —derived from Momordica cymbalaria. Preprocessing steps followed were conversion of images in to gray scale, resizing to 224 × 224 and normalizing to the scale of [0,1]. To increase the number of images for training, augmentation was used. Augmentation techniques used are rotation, flipping, contrast adjustment, gaussian noise injection, and scaling. MobileNetV2 was used as a convolution encoder for feature extraction. ResNet50 is used in ablation studies to assess architectural sensitivity. Training implements an episodic N-way-K-shot framework (with N = 4, K = 3, and Q = 2). The support (K) and query (Q) sets are sampled per episode to enable meta-learning of the features. For prototype computation, KMeans (with n = 1) was used to compute class prototypes from the embedded support set. To facilitate Mahalanobis distance-based classification, an improvement for capturing intra-class variability in SEM images, the class-wise inverse covariance matrices are simultaneously calculated. Each query embedding is compared to all class prototypes using the Mahalanobis distance, and the class with the minimum distance is predicted. Probabilities are computed via softmax on the negative distance values, and predictions are compared to ground-truth labels to calculate episodic cross-entropy loss. Last stage, involves comprehensive evaluation and validation using Leave-One-Out Cross-Validation (LOOCV) to handle the limited dataset size. Performance is analysed using classification reports, t-SNE visualizations of the embedding space, and ablation studies.

Fig. 1.

System architecture of few-shot learning for SEM image classification of phyto-synthesized nanoparticles.

Table 1, shows the visual illustration of an N-way-K-shot task (N = 4, K = 3, Q = 2) constructed from SEM images of phyto-synthesized nanoparticles. Each episode includes a support set of 3 images per class and a query set of 2 images per class for evaluation.

Table 1.

An N-way-K-shot classification task for few-shot learning of SEM images.

	Training task
	Support set
K = 3	AgNO3 root	AgNO3 fruit	CaCO3 root	CaCO3 fruit
	Query set
K = 2
	N = 4

Training configurations for model performance

Evaluated three distinct training configurations to benchmark model performance under low-data regimes: (1) Baseline Transfer Learning, (2) Euclidean Prototype Averaging, and (3) Proposed Episodic Few-Shot Learning with Mahalanobis distance.

• i.Configuration 1 – Baseline Transfer Learning: A conventional transfer learning pipeline was employed using MobileNetV2 and ResNet50 as feature extractors. The final classification was performed using a softmax layer trained on a small number of images. This served as a baseline to assess the performance of pre-trained CNNs on limited data without few-shot learning mechanisms.
• ii.Configuration 2 – Prototype Averaging with Euclidean Distance: A metric-based few-shot learning approach was applied. Feature embeddings were extracted using a pre-trained CNN encoder, and class prototypes were computed as the mean of support set embeddings. Classification was done using Euclidean distance between query embeddings and class prototypes. This approach introduces prototype-based reasoning but does not incorporate class-wise variance.
• iii.Configuration 3 – Proposed Method (Prototype + Mahalanobis + KMeans + Episodic Training): This is the full few-shot learning framework adopted in this study. Episodes were constructed using the N-way-K-shot protocol (N = 4, K = 3, Q = 2). Class prototypes were computed using KMeans clustering (K = 1) over support embeddings. Mahalanobis distance was used to account for class-specific distribution by incorporating inverse covariance matrices. The encoder was trained episodically, and data augmentation was applied to enhance generalization. This setup yielded the highest classification performance and is the central contribution of this work.

The training was carried out for 150 epochs for Configuration 3, 200 Epochs for Configuration 1 and 2, with 5 episodes generated per epoch to simulate varied classification tasks. The encoder used for feature extraction was either MobileNetV2 or ResNet50, both pretrained on ImageNet. For fine-tuning, only the last five layers of the base CNN were made trainable in Configuration 1, whereas the last twenty layers were unfrozen and fine-tuned in both Configuration 2 and Configuration 3. A dropout layer with a rate of 0.3 was applied to reduce overfitting, followed by batch normalization for improved training stability. The model was trained using the Adam optimizer with a learning rate of 0.0001. While batch size varied dynamically per episode based on the number of support and query images, it typically involved 20 images per task (4 classes × (3 support + 2 query)). The loss was computed using categorical cross-entropy on the negative distance scores between query embeddings and class prototypes. For evaluation, Leave-One-Out Cross-Validation (LOOCV) was employed.

Results

Characterization of calcium and silver nanoparticle

The SEM analysis of synthesized nanoparticles provides valuable insights into their shape and size. The calcium nanoparticles exhibited spherical structures clustered in clumps, with sizes ranging from 200 to 600 nm (Fig. 2a and b). This observations are consistent with previous where the nucleation and spherical growth of CaCO₃ nano and microparticles has been reported²⁸.The speed, time and temperature at which the nanoparticles are stirred during their synthesis plays crucial role in final size and aggregation properties. A research on these parameters effecting the size of calcium carbonate nanoparticles has been reported, wherein increasing the mixing time and speed have had profound effect on the reduction in particle size²⁹.

Fig. 2.

(a) SEM imaging of calcium nanoparticles produced from fruit and root samples respectively. (b) SEM imaging of silver nanoparticles produced from fruit and root samples respectively.

In contrast, the silver nanoparticles synthesized using M. cymbalaria extracts showed no distinct or consistent morphology, with sizes unusually ranging from 3 to 25 μm in both root- and fruit derived samples. Such large and irregular structures are indicative of aggregation or poorly regulated synthesis, possibly due to inadequate reactions within metabolites in plant extract. The aggregation observed is also reported due to insufficient stabilization and interactions between particles during drying³⁰.

Additionally, it has been reviewed that green synthesis of AgNPs is highly sensitive to reaction parameters, and lack of precise control can result in particle instability, irregular morphology, and poor dispersion.³¹.

In the present research, the irregular shapes and larger size of particles suggested that the phytochemical profile of M. cymbalaria may not have provided sufficient reducing or stabilizing capability under the conditions used for their synthesis. In contrast, parameters such as concentration of silver nitrate solution and plant extract, time, pH, and temperature were investigated and optimized for maximal yields, regulated size, and stability of silver nanoparticle in Eucalyptus camaldulensis and Terminalia arjuna extracts³².

The methanolic plant extract was used for the preparation of calcium and silver nanoparticles, wherein the change in colour during the reaction confirmed the formation of nanoparticles. To further validate, both calcium and silver nanoparticles were subjected to UV spectroscopic, EDAX and FTIR analysis. The UV spectroscopic analysis of calcium nanoparticles was recorded in the range of 300–800 nm and common peak positions were observed between 300 nm and 400 nm (Fig. 3a and b). The distinctive peak at 650 nm in both root and fruit samples indicated the presence of compounds such as Saponins, flavonoids, or phenolic which are often associated with bioactive properties. The similarity of the peaks in both the samples implies a shared chemical composition or structural feature responsible for this absorption. While there are prior reports on calcium oxide being used for nanoparticle synthesis from plant extracts^33,34 fewer reports are available for calcium carbonate based green synthesis³⁵. In either case, the peaks observed are in the range of 250–350 nm.

Fig. 3.

(a) UV spectroscopy graph for calcium nanoparticles produced from fruit and root samples respectively. (b) UV Spectroscopy graph for silver nanoparticles produced from fruit and root samples respectively.

Silver nanoparticles exhibited a characteristic absorption peak in the range of 500–700 nm in the root and fruit samples. Notably, the root extract showed a pronounced increase in absorbance within this region, corresponding to the expected surface plasmon resonance (SPR) band of silver nanoparticles. Previously, uv-visible spectrophotometric analysis of AgNPs was recorded over the wavelength range of 350–750 nm to capture their characteristic optical behaviour^36,37.

The energy-dispersive X-ray spectroscopy (EDX or EDS) was used for quantitative and qualitative detection of elements present in the sample. The EDX spectrum analysis graph of both fruit and root samples reveals the increasing composition of Ca, C and O in nanoparticles, along with other elements such as P and Silicon in very minute quantities .Oxygen is dominant in both the samples, contributing 42–45% by weight and 46–50% by atomic percentage, followed by Calcium, accounting for 30% by weight and 13% by atomic percentage. The presence of carbon is also significantly higher, with 20–25% by weight and 35–40% by atomic percentage (Table 2a). These findings, especially the percentage of Calcium are in consistent with green-synthesized CaNPs from Acacia arabica³⁸. Trace elements, including Silicon and Phosphorus, are detected in the fruit sample at minimal weight percentages of 0.30% and 0.40%, respectively. These results suggest that both samples are rich in oxygenated and calcium-containing compounds, with a substantial carbonaceous component. Similar elements including C, O, Mg, Si, and Ca with their respective atomic and percentage weights were found in the EDX spectrum of CaO nanoparticles synthesized from leaf extract of Morus nigra³⁹.

Table 2.

(a) EDAX analysis of the calcium nanoparticles produced from fruit and root samples. (b) EDAX analysis of the silver nanoparticles produced from fruit and root samples.

Elements	Atomic %		Weight %
	Fruit	Root	Fruit	Root
(a)
C	39.17	35.23	26.42	23.91
O	46.67	51.89	41.94	46.91
Si	0.19		0.3
P	0.23		0.4
Ca	13.75	12.88	30.94	29.17
(b)
C	60.5	63.25	44.5	49.45
N	3.99		3.42
O	32.44	34.7	31.79	36.13
Ag	3.07	2.05	20.29	14.41

In silver nanoparticles sample, the presence of Carbon (C), Oxygen (O), Silver (Ag), and Nitrogen (N) was noted by the analysis. Carbon is the most abundant element in both samples, with 45–50% by weight, indicating a dominant organic component. Oxygen is also prominent, contributing 32–36% by weight, reflecting oxidation or organic functional groups. Silver nanoparticles are successfully incorporated, as evidenced by their distinct peaks in the spectrum, with a higher content in the fruit sample at 15–20% by weight (Table 2b). These results confirm the presence of silver nanoparticles within a matrix primarily composed of carbonaceous and oxygen-rich materials.

Additionally, nitrogen is detected at 3.42% by weight in the fruit sample, potentially originating from organic molecules or synthesis residues. This is in contrast to a report published on synthesis of silver nanoparticles from a terrestrial fern, Gleichenia Pectinat⁴⁰. The EDX spectrum of the fern-derived silver nanoparticles indicated that oxygen was present in the highest proportion, followed by silicon, with no detec carbon content. A similar report on elemental patterns of silver nanoparticles synthesized from Neem and turmeric was observed⁴¹, highlighting the role of organic molecules in capping and reducing Ag⁺ ions during nanoparticle formation.

The FTIR spectrum of the synthesized Calcium nanoparticles shows successful presence of functional groups (Fig. 4a and b). The carbonate ion exhibits distinct vibrational features in the infrared spectrum. Strong symmetric stretching vibrations were observed in the range of 1400–1500 cm⁻¹, while asymmetric stretching modes appeared around 1000–1100 cm⁻¹, giving rise to sharp absorption peaks. Additionally, in-plane C–H bending vibrations occurred in the 600–700 cm⁻¹ region, whereas out-of-plane (arc) C–H bending vibrations were typically detected between 800 and 900 cm⁻¹. In the material, no peaks in the longer wavelength range or unwanted bonds was detected. The minor presence of O-H peaks shows that the structure of CaCO₃ nanoparticles does not depend highly on the presence of methanol used during the extraction process. Peaks at 700–600 cm⁻¹ and 800–900 cm⁻¹ corresponding to C–H bending vibrations further supported the presence of well-defined carbonate groups⁴² The absence of major peaks beyond these regions suggests minimal contamination or residual solvents, and the weak O–H signals indicate that the synthesis was not significantly influenced by methanol⁴³. Similar FTIR patterns confirming the purity and stability of plant-mediated CaO nanoparticles have been reported⁴⁴.

Fig. 4.

(a) FTIR spectrum for calcium nanoparticles produced from fruit and root samples respectively. (b) FTIR spectrum for silver nanoparticles produced from fruit and root samples respectively.

The FTIR spectrum of silver nanoparticles shows a number of absorption peaks that indicate the presence of functional groups. The broad peak at 3276–3278 cm⁻¹ indicates O-H or N-H stretching vibrations, which indicates the alcohols, phenols, or amines. The peaks at 2977 –2849 cm⁻¹ show C-H stretching, common characteristic of alkanes. A prominent peak located at 1739–1740 cm⁻¹ is associated with C = O stretching vibrations, which is indicative of the carbonyl group found in compounds like esters, aldehydes, or ketones. A peak in the vicinity of 1625–1630 cm⁻¹ implies C = C stretching, typically observed in alkenes or aromatic substances. Notable peaks at 1515–1516 cm⁻¹ and 1462–1463 cm⁻¹ are linked to the stretching of aromatic C = C bonds or the bending vibrations of CH₂. Additional absorption bands observed at 1377 cm⁻¹ are attributed to CH₃ bending vibrations, while bands in the 1156–1029 cm⁻¹ region correspond to C–O stretching or bending modes characteristic of esters or ethers. Together, these features provide clearer insight into the chemical composition of the sample. Low frequency peaks at 529 cm⁻¹ and 409 cm⁻¹ may correspond to bending vibrations involving heavier atoms or metal-ligand bonds. These findings align with the results of several green synthesised silver nanoparticles including a recent finding⁴⁵, who reported similar biomolecular signatures in FTIR spectra of biosynthesized AgNPs, confirming successful capping by phytochemicals.

Experimental results for classification of SEM images

This section discusses the results obtained for classification of SEM images by the proposed method. Table 3, summarizes the quantitative results obtained for all the three configurations, where a consistent improvement is observed from Configuration 1 to Configuration 3. Among the three configurations, Configuration 3 (Proposed Method) demonstrated the highest performance, with MobileNetV2 achieving 95% Accuracy, significantly surpassing both its own earlier configurations and the corresponding ResNet50 results. MobileNetV2 achieved Precision of 96%, Recall of 95% and F1-Score of 95% under Configuration 3. In contrast, Configuration 1 (Baseline Transfer Learning) showed the lowest scores, 50% of accuracy, precision, recall and 49% F1-score, highlighting the limitations of standard fine-tuning in low-data regimes. The Configuration 2 (Prototype Averaging with Euclidean Distance) offered moderate improvements, 65% Accuracy, 66% Precision, 65% Recall and 64% F1-score.

Table 3.

Quantitative results of all three configurations.

	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)
Configuration 1
MobileNetV2	50	50	50	49
ResNet50	45	40	45	42
Configuration 2
MobileNetV2	65	66	65	64
ResNet50	55	56	55	53
Configuration 3
MobileNetV2	95	96	95	95
ResNet50	60	58	60	58

Trained MobileNetV2 (Configuration 3) model had a compact size of 8.98 MB, which is suited for deployment in resource-constrained environments. The ResNet50 model, at 90.57 MB, appears overparameterized for the few-shot SEM classification task An ablation study was performed to assess the contribution of each component in the proposed few-shot learning pipeline. Table 4, provides the summary of the ablation study. Removing the Mahalanobis distance led to a notable drop in accuracy (from 85% to 70%), indicating its critical role in modeling class separability. Interestingly, excluding data augmentation improved performance to 95%, suggesting that in a limited data regime, augmentation may introduce undesirable variability. Due to the inherent variance in N-way–K-shot episodic sampling, performance variations across configurations reflect practical drifts rather than guaranteed statistically significant improvements.

Table 4.

Ablation study summary.

Setup	Accuracy (%)	Interpretation
Full setup	85	With KMeans + Mahalanobis + Augmentation + Episodic Training
No KMeans	80	KMeans contributes modestly to refining prototypes
No Mahalanobis	70	Mahalanobis distance is important for your data’s feature distribution
No Episodic Training	85	Suggests encoder pretraining is sufficient; episodic training may be optional
No Augmentation	95	Better: Augmentation might be introducing noisy variations due to few-shot setup

t-SNE (t-distributed Stochastic Neighbor Embedding) plots for MobileNetV2 are shown in Fig. 5. t-SNE visualizations using MobileNetV2 as the encoder highlight the progression in embedding quality across the three configurations. In Configuration 1, the class-wise clusters are poorly formed with overlap, suggesting limited discriminative ability. Configuration 2 shows better separation, with clearer boundaries emerging between classes. Configuration 3, exhibits comparatively compact and better-separated clusters. These plots suggest that the combined use of episodic training and distance-aware embedding strategies contributes to improved feature organization in the few-shot learning. However, the t-SNE plots are intended as qualitative illustrations rather than statistical evidence of significance.

Fig. 5.

t-SNE plots for MobileNetV2 for three configurations.

Ablation study and t-SNE visualization analyses are intended to provide only component-level interpretability. Overall, the ablation study and t-SNE visualization analyses are intended to provide component-level interpretability within the proposed framework, while acknowledging the variability inherent to few-shot SEM image classification.

Confusion matrices, Fig. 6, reveal fewer misclassifications in AgNP vs. CaCO₃ categories with MobileNet.

Fig. 6.

Confusion Matrix for MobileNetV2.

Figure 7 shows episodic training loss and accuracy over epoch for MobileNetV2 model of Configuration 3. The training loss for Configuration 3 using MobileNetV2 shows a consistent decline over 150 epochs, stabilizing to near zero, while the accuracy rapidly improves and converges above 95%, demonstrating convergence and generalization capability of the model under few-shot learning.

Fig. 7.

Episodic training loss and accuracy over time over epoch for Configuration 3.

A paired t-test was conducted to compare classification accuracy with and without the Mahalanobis distance metric. A paired t-test (t = − 0.6470, p = 0.5338) revealed no statistically significant difference in classification accuracy with and without the Mahalanobis distance metric. While the Mahalanobis-based configuration exhibited slightly more consistent performance in some runs, the variability across trials indicates that its impact was not consistently beneficial (Table 5).

Table 5.

Accuracy trends of t-test across different runs.

Run	With Mahalanobis Accuracy (%)	Without Mahalanobis Accuracy (%)
1	70	90
2	85	90
3	90	80
4	90	90
5	85	85
6	85	85
7	85	80
8	65	80
9	95	85
10	80	85

Limitations and future research

The dataset used comprised a small number of labelled SEM images per class (i.e., five per class), which, while aligned with the few-shot learning paradigm, may limit the generalizability of the model to broader SEM applications with higher intra-class variability. Impact of data augmentation was found to be inconsistent, indicating that more domain-specific augmentation techniques—tailored to the unique characteristics of SEM imagery is required to enhance model robustness⁴⁶. Transformer-based encoders or vision-language models, as demonstrated in the recent hybrid Vision Transformer–ResNet50 model developed for SEM image classification⁴⁷ can be looked forward as a future work. Inclusion explainable AI (XAI) techniques—such as SHapley Additive exPlanations (SHAP) or attention-based visual explanations will improve interpretability, in biomedical and materials science domain applications⁴⁸.

Conclusion

This study demonstrated the synthesis and characterization of silver and calcium carbonate nanoparticles using root and fruit extracts of M. cymbalaria, validated through UV-Visible spectroscopy, FTIR, SEM, and EDAX analyses. Though, the calcium nanoparticles in this study exhibited favourable morphology and dispersion, silver nanoparticles require further process refinement to ensure nanoscale formation. Optimizing parameters such as reaction time, precursor-to-extract ratio, and pH will be crucial in improving nanoparticle quality and ensuring suitability for biomedical applications. System architecture for SEM image classification based on Few-shot learning (FSL) has been proposed in this work. SEM image dataset had four classes (AgNO₃-root, AgNO₃-fruit, CaCO₃-root, CaCO₃-fruit) with 20 images (4 × 5 images per class) in total. Under this architecture three configurations, Baseline Transfer Learning, Euclidean Prototype Averaging, and Proposed Episodic Few-Shot Learning with Mahalanobis distance were evaluated. Configuration 3, integrating MobileNetV2 with episodic training, KMeans prototype computation, and Mahalanobis distance, achieved the classification accuracy of 95%, Precision of 96%, Recall of 95% and F1-Score of 95%. The findings establish few-shot learning as a promising approach for SEM image-based nanoparticle classification, offering a scalable pathway for analysis in scenarios constrained by limited imaging data.

Author contributions

The paper conceptualization, methodology, software and validation was carried out by 1 st author. The investigation, writing-original draft preparation and project administration, has been carried out by 2 nd author. The formal analysis, resources and data curation have been done by 3 rd, 4 th, 5 th, 6 th and 7 th authors.

Funding

Open access funding provided by B.M.S. College of Engineering.

Data availability

Sample dataset used for this study is available at the link [https://github.com/umadeviv/SEM-Images]. Complete dataset will be provided upon request to the corresponding author.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval

The research on Momordica cymbalaria was conducted in accordance with the IUCN Policy Statement on Research Involving Species. Also, the guidelines laid by Research ethical committee, Government College (A), Anantapur organization legislation was followed. The herbarium rules, field Collections of this wild Species, field numbers, voucher specimen deposition and GPS etc. is maintained in the institute. Ethical committee approval was obtained prior to collection and research on the plant.