Page 900
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue I, January 2026
Deep Learning-Based Detection and Classification of Dental
Conditions
Mrs. Priyata Mishra, Jaikumar Dewangan, Avijit Agrawal, Bhupendra Dewangan
Dept. of Computer Science and Engineering SSIPMT, Raipur, C.G., India
DOI:
https://doi.org/10.51583/IJLTEMAS.2026.150100078
Received: 28 January 2026; Accepted: 02 February 2026; Published: 10 February 2026
ABSTRACT
Dental conditions, particularly calculus and caries, are among the primary causes of poor oral health. They
frequently result in pain, discomfort, infections, and occasionally even tooth loss. To prevent these problems
from getting worse, early detection is vital. In many places, though, access to dental specialists is still restricted.
In order to automatically identify and categorize four crucial aspects of oral health—caries, calculus,
discoloration, and healthy teeth—this study proposes a deep learning approach. By using readily available
intraoral images, the model removes the need for radiographic methods up to an extent, allowing for affordable
and non-invasive screening. The system uses Transfer Learning with a ResNet-based Convolutional Neural
Network (CNN), which provides excellent feature extraction and quick learning even with smaller datasets. We
trained the model extensively with pre-processed intraoral images and fine-tuned the higher convolutional layers.
In tests on the validation dataset, the model reached a classification accuracy of 97.35 percent. This shows it can
accurately tell the difference between healthy and diseased dental states with great precision and low variance
0.34 percent across the cross-validation tests. We can expand the community's access to early detection through
integrating them into applications. This framework marks a major breakthrough in the use of deep learning
models in preventive dentistry with its rapid, scalable, and accurate screening that improves patient care and
public health outcomes.
Keywords; Dental Conditions, Deep Learning, CNN, AI.
INTRODUCTION
Regardless of age or location, dental caries, plaque, and calculus affect a significant portion of the population.
This makes oral diseases some of the most common illnesses [1]. Despite being mostly preventable, these
conditions often remain untreated due to a lack of early detection and insufficient access to dental care facilities,
particularly in developing and rural areas [2]. Poor oral health can lead to pain and tooth decay, and it’s also
connected to other diseases like diabetes and heart disease. This highlights the importance of early detection [3].
Effective preventative care also depends on consistently monitoring healthy teeth. Most dental exams rely on
clinical inspections and if having more symptoms goes to radiographic imaging, both require trained specialists,
equipment, and in-person visits. This reliance creates significant barriers to quick diagnoses, as many people
delay care until they feel discomfort, leading to costly and invasive treatments later. In addition, regional and
socioeconomic differences in dental resources increase the gap in access to oral healthcare [4]. We need to create
affordable, accessible diagnostic tools that can be used outside of traditional clinical settings. Recent
advancements in Artificial Intelligence (AI), mainly Deep Learning (DL), have greatly improved medical image
analysis. Among Deep Learning methods, Convolutional Neural Networks (CNNs) stand out for their
exceptional ability to automatically extract spatial and visual features from complex images [5]. Architectures
such as ResNet allow for deeper networks to be trained without losing performance, increasing the accuracy of
visual interpretation without performance degradation [6]. Additionally, Transfer Learning, where a model
trained on huge dataset is adjusted for a specific area, can significantly improve performance, even with limited
medical data, speeding up development [7]. This research uses CNN-based Transfer Learning to identify and
classify four major dental categories—caries, calculus, discoloration, and healthy teeth—from non-radiographic,
intraoral images. The focus on real-image data, which can be captured by simple cameras or smartphones,
Page 901
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue I, January 2026
enhances the system’s applicability for community use and improves accessibility for remote dental assessments.
Including the 'healthy teeth' category is important for building a foundation for preventive oral health monitoring.
With a classification accuracy of 97.35%, the proposed model shows great promise for real-world use as a digital
diagnostic tool for broad preventive oral screening.
LITERATURE REVIEW
The field of dental diagnostics has witnessed significant evolution, driven by the need for AI-based image
analysis systems capable of augmenting expert-level visual interpretation. Historically, dentists have relied on
manual inspection and radiographic imaging for assessing caries and other oral abnormalities. Although
clinically effective, these methods suffer from inherent subjectivity and limited scalability for mass screening
purposes [8].
Evolution of Computer Vision in Dental Diagnostics
Several initial studies explored applying conventional computer vision methods in dentistry. Earlier, the image
processing techniques focused on thresholding and edge detection for caries detection; however, these
approaches lacked robustness against variations in illumination, texture, and angle typical of real-world intraoral
images [9]. The subsequent introduction of classical machine learning algorithms, particularly Random Forest
and Support Vector Machines (SVMs) classifiers, improved the detection of dental calculus and enamel
anomalies but still necessitated labour-intensive, manual feature extraction, limiting their potential for wide-
scale, automated systems [10].
The Deep Learning Paradigm and Architectural Advance
Deep Learning (DL) marked a fundamental paradigm shift. CNNs have demonstrated the capability to achieve
human-level accuracy in classifying numerous medical conditions from images—including skin lesions, fundus
images for diabetic retinopathy, and dental pathologies visible on radiographs [11]. Crucially, CNNs can
automatically learn intricate, multi-level features directly from raw image pixels without the need for domain-
specific feature engineering, a vital factor for managing the high variability of intraoral images captured under
non-clinical conditions.He et al. (2015) introduced the Residual Network (ResNet), a deep CNN architecture
that addresses the vanishing gradient problem through the use of skip connections, making it possible to train
networks with hundreds of layers effectively [6]. This innovation is credited for the architecture itself [12], which
has yielded breakthrough performance in complex image classification tasks by enabling the network to learn
residual mapping.
Table 1. Dataset Distribution
Category
Number of Images
Percentage
Caries
2,381
33.92%
Calculus
1,307
18.62%
Healthy Teeth
1,501
21.38%
Discoloration
1,831
26.08%
Total
7,020
100%
Transfer Learning for Domain Specialization
Transfer Learning has since extended these architectural benefits into healthcare, particularly where high-quality,
domain-specific datasets are limited. By initializing models with ImageNet-trained weights and then fine-tuning
on dental datasets, the model inherits powerful low-level visual features (e.g., edges, corners) learned from
Page 902
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue I, January 2026
millions of general images. This process significantly accelerates convergence and enhances To ensure
consistent input quality and increase model robustness, a structured preprocessing pipeline was developed. In
order to maintain equal input size, each of the image dimensions was resized to 224x224. To normalize the
distribution of the data to simplify further processing and accelerate convergence, we normalized pixel intensity
by normalizing it using a division by 255 in [0, 1].
To address class imbalance and enhance model generalization, a selective class-adaptive data augmentation
strategy was used. To create the illusion of mirrored orientations, all images were subjected to horizontal
flipping. Strong augmentation (e.g., ±15%–40% brightness, 0.7–1.4X contrast adjustment) was used for the
minority class (Calculus). The majority of classes received light augmentation with a ±15% brightness variation
limit. This adaptive approach ensured a more balanced representation and reduced overfitting.
Fig. 1. Flowchart of Proposed Methodology
Page 903
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue I, January 2026
Contemporary AI Applications in Dental Screening
Recent dental research directly supports the application of advanced CNNs and Transfer Learning using non-
radiographic imagery. Studies leveraging ResNet50 and similar architectures have reported high accuracy in
detecting caries and calculus using intraoral photographs [14]. A few pioneering works have attempted to classify
healthy versus diseased teeth, showing strong potential for self-assessment applications via smartphone cameras
[15]. This trend indicates a growing recognition that surface-level pathologies can be reliably identified without
X-rays, paving the way for ubiquitous screening tools. Building upon this foundation, the present study integrates
these architectural and methodological advances into a robust, multi-class system specifically targeting the
foundational conditions of caries, calculus, discoloration, and the crucial 'healthy teeth' status from real intraoral
images. This approach simultaneously addresses the clinical need for high diagnostic accuracy and the public
health imperative for improved accessibility.
METHODOLOGY
The methodology details the refined, deep learning-based approach specifically designed for the automated and
accurate classification of the four core dental health states using solely intraoral images. This system is
engineered to provide an accessible diagnostic alternative to traditional clinical inspections.
Data Acquisition and Preprocessing
The study made use of a combined dataset of 7,020 intraoral dental images in JPG format that were obtained
from Mendeley Data and Kaggle, two publicly accessible and peer-reviewed repositories [16] [17]. All photos
were meticulously and manually annotated by skilled dental professionals to guarantee high-quality labels.
To ensure consistent input quality and increase model robustness, a structured preprocessing pipeline was
developed. In order to maintain equal input size, each of the image dimensions was resized to 224x224. To
normalize the distribution of the data to simplify further processing and accelerate convergence, we normalized
pixel intensity by normalizing it using a division by 255 in [0, 1]. To address class imbalance and enhance model
generalization, a selective class-adaptive data augmentation strategy was used. To create the illusion of mirrored
orientations, all images were subjected to horizontal flipping. Strong augmentation (e.g., ±15%–40% brightness,
0.7–1.4X contrast adjustment) was used for the minority class (Calculus). The majority of classes received light
augmentation with a ±15% brightness variation limit. This adaptive approach ensured a more balanced
representation and reduced overfitting.
Model Architecture and Training Strategy
The proposed system uses a transfer learning technique with ResNet50 architecture [7]. The architecture is
composed of two primary parts: a feature extraction backbone, and a custom classification head.
Table 2. Model Summary
Layer (type)
Output
Shape
Param #
input_layer_9 (InputLayer)
(None, 224,
224, 3)
0
get_item_12 (GetItem)
(None, 224,
224)
0
get_item_13 (GetItem)
(None, 224,
224)
0
get_item_14 (GetItem)
(None, 224,
224)
0
stack_4 (Stack)
(None, 224,
224, 3)
0
Page 904
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue I, January 2026
add_4 (Add)
(None, 224,
224, 3)
0
resnet50 (Functional)
(None, 7, 7,
2048)
23,587,712
global_average_pooling2d
(GlobalAveragePooling2D)
(None, 2048)
0
dense_8 (Dense)
(None, 256)
524,544
dropout_4 (Dropout)
(None, 256)
0
dense_9 (Dense)
(None, 4)
1,028
Fig. 2. Model Summary
Model Initialization and Custom Head
The backbone was initialized using ResNet50 weights pretrained on ImageNet and set as a non-trainable feature
extractor in the first stage. A custom classification head was added comprising a Global Average Pooling 2D
layer for spatial dimensions reduction, a ReLU activation function, L2 regularization (λ = 1e−4), a Fully
Connected (Dense) layer with 256 units, a Dropout layer (rate = 0.5) for regularization, and an Output layer
with 4 neurons and Softmax activation for multi-class classification. In order to address class imbalance, the
Scikit-learn “balanced” class weighting approach was used during training.
Training Methodology: A two phase transfer learning method was used to enable good adaptability of features.
Phase 1 — Feature Extraction: All the convolutional layers of ResNet50 were pre-trained, while only the custom
head was trained over 25 epochs by freezing ResNet Base as feature extractor with Adam optimizer (learning
rate =1e−4) and Sparse Categorical Cross-Entropy loss. This step allowed the classifier to learn how to map
pre-trained generic features into dental disease classes.
Phase 2 — Fine-Tuning: The ResNet50 backbone's initial 100 layers were set to a frozen state. A very low
learning rate (LR=1e-5) was used to fine-tune the entire model for 12 additional epochs. This controlled process
avoids catastrophic forgetting and allows the network to adapt its high-level features to the specific visual
characteristics of the intraoral images. Early stopping and learning rate reduction callbacks were utilized to
ensure stable convergence and prevent overfitting.
Evaluation Metrics and Validation
A 3-fold cross-validation scheme was used to make sure each image was used as a test sample once in order to
guarantee a reliable and objective assessment. The training subset was further split into an 80/20 split for training
and validation, respectively, within each fold. Accuracy, Balanced Accuracy, Macro F1-Score, Precision, and
Recall were among the metrics used to evaluate the model's performance, in addition to a thorough Confusion
Matrix analysis. In order to ensure generalization across data splits, the reported results are the average
performance across all three folds.
RESULTS AND DISCUSSION
The rigorous evaluation of the proposed ResNet50-based deep learning model, employing the two-phase
Transfer Learning strategy, yielded exceptional performance metrics, strongly validating the system's potential
for automated, non-radiographic dental screening.
Page 905
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue I, January 2026
Performance Evaluation and Model Consistency
The performance of the model was assessed through a 3-fold cross-validation approach on the 7,020 images
dataset. The summarized performance of all folds is shown in Table II. The system exhibited strong stability
across partitions, with very small variation (σ ≤ 0.34%), suggesting generalization performance more robust
rather than over-fitted to a single data split. An average overall Accuracy of 97.35% verifies the model as a very
accurate diagnostic method.
Table 3. Performance metrics
Metric
Fold 1
Fold 2
Fold 3
Mean ± STD
Test
Accuracy
97.48%
97.69%
96.88%
97.35% ±
0.34%
Balanced
Accuracy
97.65%
97.85%
97.03%
97.51% ±
0.35%
F1-Score
(Macro)
97.72%
97.90%
97.16%
97.59% ±
0.32%
Per-Class Performance and Clinical Reliability
Detailed per-class metrics (Figure 3) confirm the model's clinical utility. The model achieved near-perfect
performance for key categories: Calculus detection reached a Recall of 0.9985 (no false negatives), and the
classification of Healthy Teeth attained an F1-Score of 0.9957 (Precision 0.9960, Recall 0.9953). The high
precision and recall for caries (0.9461 and 0.9887 respectively) underscore the model's reliability in identifying
the most prevalent pathology.
Table 4. Per-class Performance Metrics
Class
Precision
Recall
F1-Score
Caries
0.9461
0.9887
0.9669
Calculus
0.9805
0.9985
0.9894
Healthy
0.9960
0.9953
0.9957
Discoloration
0.9882
0.9181
0.9519
Accuracy
—
—
0.9735
Macro Avg
0.9777
0.9751
0.9760
Weighted Avg
0.9742
0.9735
0.9733
Comparative Context and Architectural Impact
The achieved accuracy of 97.35% represents a substantial advancement over recent literature in dental AI. Our
framework demonstrates a notable 5.5-7.6% improvement in accuracy for four-class classification. This superior
Page 906
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue I, January 2026
performance is directly attributable to the combined benefits of the ResNet50 architecture and the optimized
two-phase Transfer Learning strategy.
Error Analysis and Implications
Analysis of the aggregated Confusion Matrix (Figure. 3 - Normalized Confusion Matrix) revealed a critical
finding: approximately 79.6% of all misclassifications occurred between the Discoloration and Calculus classes.
This 2.65% inter-class confusion rate is clinically understandable, as early calculus deposits and certain
pigmentation patterns share visual similarities under varied illumination. This error is manageable given the high
overall accuracy, but it provides a clear direction for future refinement, such as utilizing more advanced colour
normalization techniques or domain-specific loss functions.
Fig. 4. Confusion Matrix
CONCLUSION
In this paper, we have demonstrated high performance and feasibility of a deep-learning based approach to
automatic classification of four basic oral-health states (caries, calculus, discoloration, healthy teeth) employing
only readily obtainable intraoral images. The model obtained a remarkable average classification accuracy of
97.35% on the test set, when by using Transfer Learning based ResNet architecture. While the high diagnostic
precision and use of prevalent RGB images emphasize how AI can support decision making in clinician settings,
chances to expand dependable dental screening into non-clinician environments are feasible.
The correct distinction of the 'healthy teeth' state is especially important, since it forms a basis for efficient
preventive treatments. This mechanism has low implementation cost and is generally available, supporting the
proactive management of oral health. Next, through collaboration with mobile platforms, large-scale
community-based dental wellness screening programs will be facilitated and early disease diagnose and oral care
education can benefit people from all over the world.
REFERENCES
1. A. Qayyum et al., “Dental caries detection using a semi-supervised learning approach,” Scientific Reports,
vol. 13, no. 1, 0 2023, doi: 10.1038/s41598-023-27808-9. [Online]. Available:
https://pubmed.ncbi.nlm.nih.gov/36639724/
Page 907
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue I, January 2026
2. M. E. Northridge, A. Kumar, and R. Kaur, “Disparities in Access to Oral Health Care.,” Annual Review of
Public Health, vol. 41, no. 1, pp. 513–535, 0 2020, doi: 10.1146/annurev-publhealth-040119-094318.
[Online]. Available:
https://pubmed.ncbi.nlm.nih.gov/31900100/
3. M. S. Lipsky, T. Singh, G. Zakeri, and M. Hung, “Oral Health and Older Adults: A Narrative Review.,”
Dentistry Journal, vol. 12, no. 2, p. 30, 0 2024, doi: 10.3390/dj12020030. [Online]. Available:
https://www.mdpi.com/2304-6767/12/2/30
4. K. Griffith, J. Bor, and L. Evans, “The Affordable Care Act Reduced Socioeconomic Disparities in Health
Care Access.,” Health Affairs, vol. 36, no. 8, pp. 1503–1510, 0 2017, doi: 10.1377/hlthaff.2017.0083.
[Online]. Available:
https://www.researchgate.net/publication/318733499_The_Affordable_Care_Act_Reduced_Socioecono
mic_Disparities_In_Health_Care_Access
5. Y. Xu, R. Quan, W. Xu, Y. Huang, X. Chen, and F. Liu, “Advances in Medical Image Segmentation: A
Comprehensive Review of Traditional, Deep Learning and Hybrid Approaches.,” Bioengineering (Basel,
Switzerland), vol. 11, no. 10, p. 1034, 0 2024, doi: 10.3390/bioengineering11101034. [Online]. Available:
https://www.mdpi.com/2306-5354/11/10/1034
6. K. He, X. Zhang, S. Ren, and J. Sun, "Deep Residual Learning for Image Recognition," Proceedings of
the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016.
[Online]. Available:
https://openaccess.thecvf.com/content_cvpr_2016/html/He_Deep_Residual_Learning_CVPR_2016_pap
er.html
7. M. Raghu et al., "Transfusion: Understanding Transfer Learning for Medical Imaging," Advances in
Neural Information Processing Systems (NeurIPS), vol. 32, 2019. [Online]. Available:
https://papers.nips.cc/paper_files/paper/2019/hash/eb1e78328c46506b46a4ac4a1e378b91-Abstract.html
8. Schwendicke, F et al. “Artificial Intelligence for Caries Detection: Value of Data and
Information.” Journal of dental research vol. 101,11 (2022): 1350-1356.
doi:10.1177/00220345221113756. [Online]. Available: https://pubmed.ncbi.nlm.nih.gov/35996332/
9. A. M. Shervedani, H. Khodadadi, and S. I. Mousavian, "Development a computer-aided diagnosis system
for dental caries detection applying radiographic images," Computers in Biology and Medicine, vol. 196,
Part C, p. 110966, 2025, doi: 10.1016/j.compbiomed.2025.110966. [Online]. Available:
https://www.sciencedirect.com/science/article/pii/S0010482525013186
10. Choi J. Comparative effectiveness research in observational studies. J Periodontal Implant Sci. 2018
Dec;48(6):335-336. https://doi.org/10.5051/jpis.2018.48.6.335. [Online]. Available:
https://jpis.org/DOIx.php?id=10.5051/jpis.2018.48.6.335
11. A. Esteva et al., "Dermatologist-level classification of skin cancer with deep neural networks," Nature,
vol. 542, no. 7639, pp. 115–118, 2017. [Online]. Available: https://www.nature.com/articles/nature21056
12. Y. Chen et al., "Deep Transfer Learning for Multi-Pathology Classification in Dental Radiographs," IEEE
Access, vol. 7, pp. 107797–107806, 2019. [Online]. Available:
https://ieeexplore.ieee.org/document/8713964
13. Mohammad-Rahimi, Hossein et al. “Deep learning for caries detection: A systematic review.” Journal of
dentistry vol. 122 (2022): 104115. doi:10.1016/j.jdent.2022.104115 [Online]. Available:
https://pubmed.ncbi.nlm.nih.gov/35367318/
14. Kang S, Shon B, Park EY, Jeong S, Kim EK. Diagnostic accuracy of dental caries detection using ensemble
techniques in deep learning with intraoral camera images. PLoS One. 2024 Sep 6;19(9):e0310004. doi:
10.1371/journal.pone.0310004. PMID: 39241044; PMCID: PMC11379315. [Online]. Available:
https://pmc.ncbi.nlm.nih.gov/articles/PMC11379315/
15. Takahama, Ricardo Kenji et al. “Accuracy of smartphone photographs for detecting active carious lesions
in orthodontic patients.” Brazilian oral research vol. 39 e069. 8 Sep. 2025, doi:10.1590/1807-3107bor-
2025.vol39.069 [Online]. Available: https://pmc.ncbi.nlm.nih.gov/articles/PMC12419182/
16. S. Salman, “Oral Diseases Dataset,” Kaggle, 2023. Accessed: Jan. 2025. Available: [Online]. Available:
https://www.kaggle.com/datasets/salmansajid05/oral-diseases
17. "Dental Health Image Collection," Mendeley Data, 2023. [Online]. Available:
https://data.mendeley.com/datasets/