ASSIGNMENT
Iqra Basheer
Roll No: 25015919-014
Course: Data Analytics
Instructor: Dr. Omer
Department of Computer Science
University Of Gujrat
IMPACT OF GAN-CNN FRAMEWORK ON ORAL CAVITY TREATMENT
ABSTRACT
This research explores the combined use of Convolutional Neural Networks (CNNs) and
Generative Adversarial Networks for improving the diagnosis and treatment planning of oral
cavity cancers. The study utilizes the Histopathological Imaging Database for Oral Cancer
Analysis obtained from Mendeley Data to evaluate how advanced deep learning techniques can
support early cancer detection and accurate tissue classification.
In our approach, CNN models are employed to automatically analyze and classify
histopathological images, while GANs are integrated to generate synthetic samples for data
augmentation. This helps mitigate class imbalance and enhances the overall robustness of the
training dataset. By leveraging the strengths of both architectures, the proposed framework
improves learning efficiency and classification performance.
Experimental results demonstrate that the hybrid CNN–GAN model achieves an overall
classification accuracy of 94.7% in distinguishing normal oral tissue from malignant samples.
Additionally, the model shows strong diagnostic reliability, with a sensitivity of 92.3% and a
specificity of These findings indicate that the integration of GAN-based data generation with
CNN-based feature extraction significantly enhances model performance.
Overall, this study contributes to the field of Ai base assisted oncology by presenting an effective
and scalable framework for oral cancer detection. The proposed method has the potential to assist
clinicians in clinical decision-making, enable earlier diagnosis, and ultimately improve patient
outcomes.
BACKGROUND
Oral cavity cancer remains a major global health concern, with an estimated 377,713 new cases
and 177,757 related deaths reported each year worldwide. Despite advances in medical research
and treatment, survival outcomes remain poor, particularly when the disease is diagnosed at an
advanced stage. The five-year survival rate for late-stage oral cancer is approximately 50%,
highlighting the urgent need for earlier and more accurate diagnostic approaches.
Current diagnostic practices primarily depend on clinical visual examination followed by
histopathological assessment of tissue samples. While histopathology is considered the gold
standard for diagnosis, the process is labor-intensive, time-consuming, and highly dependent on
the expertise of pathologists. Moreover, subjective interpretation can lead to inter-observer
variability, potentially affecting diagnostic consistency and treatment decisions. These limitations
underscore the need for automated, reliable, and scalable diagnostic tools to support clinicians in
the early detection of oral cavity cancer.
The rapid advancement of artificial intelligence (AI), particularly in the field of deep learning,
presents powerful opportunities to address the shortcomings of traditional oral cancer diagnostic
methods. Deep learning models are capable of automatically extracting complex patterns from
medical images, reducing reliance on manual interpretation and improving diagnostic consistency.
Among these techniques, Convolutional Neural Networks (CNNs) have shown remarkable
success in medical image analysis due to their ability to learn discriminative spatial features
directly from raw image data. However, the performance of CNNs is often constrained by limited
and imbalanced datasets, which are common challenges in medical imaging. Generative
Adversarial Networks (GANs) offer an effective solution by generating realistic synthetic images
that enhance data diversity and improve model generalization. By integrating CNNs with GAN-
based data augmentation, this research is motivated to develop a more robust, accurate, and reliable
framework for oral cavity cancer detection.
OBJECTIVES
The primary objectives of this study are as follows:
To develop a CNN-based automated classification system for the analysis of oral
histopathological images.
To design and implement Generative Adversarial Network (GAN) architectures for the
generation of high-quality synthetic images aimed at augmenting limited training datasets.
To evaluate the effectiveness of hybrid CNN–GAN models in improving oral cancer
detection and classification performance.
To compare the proposed deep learning framework with traditional machine learning
methods as well as assessments provided by expert pathologists.
SIGNIFICANCE
This research addresses the growing demand for automated, accurate, and scalable diagnostic
solutions in the field of oral oncology. By leveraging deep learning techniques, the proposed
approach has the potential to reduce diagnostic time, minimize subjectivity in histopathological
interpretation, and enhance overall diagnostic accuracy. The framework is particularly valuable
for resource-limited healthcare settings, where access to specialized pathologists may be restricted.
Ultimately, this study contributes toward improving early detection, supporting clinical decision-
making, and enhancing patient outcomes in oral cancer care.
LITERARTURE OVERVIEW
Deep Learning in Medical Imaging
Recent advancements in deep learning have significantly transformed the field of medical image
analysis across multiple specialties, including radiology, pathology, and dermatology. Deep
learning models, particularly convolutional architectures, have demonstrated superior performance
in tasks such as image classification, segmentation, and disease detection by learning complex
feature representations directly from imaging data.
In medical applications where annotated datasets are often limited, transfer learning has emerged
as an effective strategy. Pre-trained models such as ResNet, VGG, and Inception, originally
trained on large-scale datasets, have been successfully adapted to medical imaging tasks. These
models enable faster convergence, improved generalization, and enhanced performance, making
them especially suitable for histopathological image analysis and cancer detection.
CNNs in Oral Pathology
Convolutional Neural Networks (CNNs) have been widely investigated for the automated
detection and classification of oral lesions, demonstrating promising results across multiple
studies. Prior research has reported classification accuracies ranging from 85% to 93%, depending
on the selected network architectures, dataset characteristics, and preprocessing techniques.
Despite these encouraging outcomes, several challenges persist. Class imbalance within
histopathological datasets often leads to biased model learning, while variations in image
acquisition protocols, staining techniques, and magnification levels limit model generalizability.
These factors highlight the need for more robust learning frameworks capable of handling data
scarcity and heterogeneity, motivating the integration of advanced techniques such as data
augmentation and generative modeling.
GANs for Medical Data Augmentation
Generative Adversarial Networks have gained significant attention as effective techniques for
generating realistic synthetic medical images. In contrast to traditional data augmentation methods
such as rotation and flipping, GAN-based approaches are capable of learning underlying data
distributions and producing high-fidelity images that closely resemble real clinical samples.
Recent studies have shown that incorporating GAN-generated images into training datasets can
lead to performance improvements of approximately 5–15% in medical image classification tasks.
Among the various GAN architectures, Conditional GANs (cGANs) enable class-specific image
generation, while Progressive GANs facilitate the synthesis of high-resolution histopathological
images. These models have demonstrated particular effectiveness in addressing class imbalance
and enhancing model generalization, making them well suited for oral cancer histopathology
analysis.
Research Gap
Although extensive research has demonstrated the effectiveness of Convolutional Neural
Networks for oral lesion classification and Generative Adversarial Networks for medical image
augmentation, relatively few studies have investigated their combined application in the context
of oral cancer diagnosis. Existing approaches often evaluate these techniques independently,
without fully exploiting their complementary strengths.
This study addresses this gap by proposing a comprehensive hybrid CNN GAN based
framework that integrates GAN-based synthetic image generation with CNN-based classification.
By leveraging both architectures, the proposed approach aims to improve diagnostic accuracy,
enhance data diversity, and increase model robustness for oral histopathological image analysis.
METHODOLOGY
Data Description
This study utilizes the Histopathological Imaging Database for Oral Cancer Analysis obtained
from Mendeley Data, comprising a total of 1,224 histopathological images. The dataset is
organized into two subsets based on image resolution and magnification levels.
The first subset contains images captured at 100× magnification, consisting of 89 images of
normal oral epithelium and 439 images of Oral Squamous Cell Carcinoma (OSCC). The
second subset includes images acquired at 400× magnification, comprising 201 normal oral
epithelium images and 495 OSCC images. This multi-resolution structure enables the evaluation
of model performance across varying magnification scales commonly used in histopathological
analysis.
All images were obtained from hematoxylin and eosin (H&E) stained tissue slides and captured
using a Leica ICC50 HD microscope. The tissue samples were collected, prepared, and
catalogued by medical professionals from 230 patients, ensuring clinical relevance and data
reliability. The diversity in magnification levels and patient samples makes this dataset suitable
for training and evaluating deep learning models for oral cancer detection.
Data Preprocessing
Figure 1:steps for Data preprocessing
In this study, we evaluated both custom CNN architectures and transfer learning–based
models for the classification of oral histopathological images.
Base Models: Pre-trained architectures were utilized as backbone networks for feature
extraction. These models were chosen due to their proven performance in medical image analysis
and their ability to generalize well with limited datasets.
Architecture Modifications: The pre-trained networks were adapted with custom classification
heads, incorporating dropout layers to prevent overfitting and batch normalization to
stabilize and accelerate training.
Activation Functions: ReLU was used in all hidden layers to introduce non-linearity, while the
Softmax function was applied to the output layer for multi-class probability prediction.
Loss Function and Optimizer: The categorical cross-entropy loss was employed to measure
prediction error, optimized using the Adam optimizer with a scheduled learning rate starting at
0.0001 to ensure smooth convergence.
Training Strategy: To improve training efficiency and model generalization, early stopping was
applied to halt training when validation loss plateaued, and model checkpointing was used to save
the best-performing model during the training process.
Hybrid GAN–CNN Framework
To enhance classification performance and address class imbalance in oral histopathological
images, a hybrid GAN–CNN framework was developed, consisting of the following phases:
Phase 1: GAN Training
The GAN is first trained on the original dataset to learn the distribution of histopathological
images, enabling the generation of realistic synthetic samples.
Phase 2: Synthetic Image Generation
Once trained, the GAN generates synthetic images specifically for minority classes, mitigating
the impact of class imbalance in the dataset.
Phase 3: Dataset Augmentation
The real and GAN-generated images are combined to form an augmented training dataset,
providing a more balanced and diverse set of samples for CNN training.
Phase 4: CNN Training
A CNN classifier (custom or transfer learning–based) is trained on the augmented dataset to
learn discriminative features for accurate oral cancer detection.
Phase 5: Validation and Testing
Model evaluation is performed exclusively on real images, ensuring that performance metrics
reflect practical diagnostic capability. This phase includes accuracy, sensitivity, specificity, and
other relevant evaluation metrics.
Figure 2: image of GAN model
RESULTS & FINDINGS
Table 1CNN Performance
Metric Value
Overall Accuracy 89.4%
Sensitivity (Recall) 85.2%
Specificity 92.1%
F1-Score 87.6%
Training Time (GPU: RTX 3090) 4.2 hours
Table 2 CNN GAN mix result
Metric Value Improvement from Baseline CNN
Overall Accuracy 94.7% +5.3%
Sensitivity (Recall) 92.3% +7.1%
Specificity 96.1% +4.0%
F1-Score 93.8% +6.2%