Hybrid CNN Models for Multimodal Synthesis in High-Fidelity Oral Cancer Detection and Characterization

According to WHO cancer is the 2nd leading cause of death in the World [1]. Oral cancer is the 13th most common cancer globally. The Worldwide index of oral cavity cancer is approximately 377713 new cases and 177757 deaths in 2020. It is more common in men and older people as compared to women [2]. Oral Cancer is a disease which causes uncontrollable growth of cell occurring in various parts of mouth including lips, tongue, hard and soft palate cheeks, etc. It is more prominent across India and covers almost one fourth of global incidents which is around 77,000 new case and 52,000 deaths reported annually [3]. Major cause of oral cancer is Tobacco consumption which includes betel-quid chewing, smokeless tobacco, alcohol consumption, poor oral hygiene, etc. Since it is at its peak in India, early detection is the most effective phase which can have a survival rate of 80% but after metastasis the chance drops to 30% [4].

The NDB-UFES dataset used in this study is based on Brazilian patient data; we acknowledge a broader challenge in oral cancer research: many existing datasets, such as RMDS, are region-specific, with several originating from northeastern India. These areas often exhibit distinct risk factors which include high consumption of tobacco and betel quid, increased sun exposure, and potential genetic influences that may not be representative of other populations. North-east has the highest rate of cancer with the leading cancer of oral and stomach cancer, risk factors tobacco and household burning of firewood. Madhya Pradesh is in the 8th position in leading oral cancers with risk factors of tobacco and paan masala [5]. Due to lack of resources and technologies early detection is difficult. More often, this disease is detected after the cancer has already started to spread i.e. the early stage which has 5 to 6 years of survival rate which is approximately. 69.5%. However, in the later stage the rate drops to 31.6% [6]. The occurrence of oral cancer rate in United States of America is 3% whereas in India and other Asian countries it is 30%, resulting in 48,000 Americans to get affected by oral cancer every year and 8500 people die annually due to this disease [7]. Out of all the countries that has got affected by oral cancer across Asia, North America, South America and Europe, the 10 most popular countries are China, India, the United States, Indonesia, Brazil, Pakistan, Bangladesh, Russia, Japan and Mexico [8]. The main cause of having a high death rate in oral cancer is not because it’s difficult to detect and diagnose it but because it’s usually found too late when the disease has already advanced [9]. As a result, models trained on such localized data may not generalize well across different demographic or geographic contexts. Our contribution introduces a new perspective by incorporating data from a different region, thereby enhancing diversity. However, to build models with wider applicability and stronger performance across populations, future studies should prioritize collecting and validating data from a broader and more diverse range of sources.

Specialized doctors and instructors were appointed in order to detect the cancer and they perform several practices which include collecting demographic data about the patients which include basic information such as name, age, gender, weight, height, consumption of any kind of tobacco and alcohol, infected area, size of lesion, location of lesion and a part of infected tissue to test in lab and after completion of all tests the specialist come to a conclusion about the cancer. This manual operations take time and require specialized labors, equipment and the diagnosis may vary due to technician trainings [10].

Advancements in Artificial Intelligence (AI) has shown great promise in enhancing the detection and diagnosis of oral cancer. AI techniques, particularly deep learning algorithms, have demonstrated high accuracy in identifying cancerous lesions from histopathological images. These methods not only reduce the workload on physicians but also improve diagnostic precision by analysing large datasets and recognizing patterns that might missed by the human eye [10].

In this paper we aim to shed light in the advancement of AI and deep learning approach for medical diagnosis, doctors and specialists utilize both advanced and traditional methods for cancer detection using microscopic images. Hence, we aim to apply AI techniques specifically machine learning (ML) and deep learning (DL) for efficient generalization capability of the model to unique data. CNN-based model is categorized under finest DL method for learning process over varieties of datasets. Here we introduced a multimodal architecture with combined data i.e., both image data and clinical data for training purpose. Initially the histopathological images are passed through the input layer of our Custom-CNN layers for feature extraction, and the feature map is combined with clinical data and passed through the dense layer. Furthermore, there is a detailed explanation about our research and implementation categorized in several sections. Section 2 provides some collective studies which were carried out during past few years and how technology plays an important role in medical diagnosis. All these research works have motivated us in bringing off this project as it was interesting to know that several branches of science and technologies can be merged together to conclude with a lifesaving achievement. Next, we have section 3 providing a detailed description about the dataset used for this research, methodologies used and the multimodal architecture. Lastly, we have result and conclusion in section 4 which provides a detailed explanation about the progress, achievements, challenges faced and some limitations.

There are few public datasets available, here we will use NDB-UFES an oral cancer and leucoplakia dataset composed of histopathological images and patient data for our research and implementation. This public dataset includes patient demographic data along with lesion images [11]. Artificial Intelligence enables automated and precise cancer diagnosis by analyzing histopathological images to identify, classify, and predict tumour characteristics with high accuracy [12]. Few practices were conducted and various systems were designed using AI and technologies for classifying cancers [13]. Our research includes distinguishing the cancer at an early stage, a multimodal data fusion approach using deep neural network is used for the features extraction and generate a model which can predict the cancer at an early stage.

3.1 Dataset description

The efficiency of our research also depends on the dataset selected for implementation. The NDB-UFES dataset composed of histopathological images and patient data. The images are of different sizes with approximately 2048 × 1536 pixels. All total of 237 images in PNG format out of which 89 are leukoplakia with dysplasia, 57 are leukoplakia without dysplasia and 91 are OSCC images captured with an optical light microscope using 10x and 40x objective attached with a microscope camera. A hematoxylin-eosin stain is used in the histopathological slides from the biopsy of patients performed between 2010 and 2021 managed by Oral Diagnosis project of Federal University of Espírito Santo (NDB-UFES) [11]. The dataset contains sociodemographic data including age, gender and skin color as well as clinical data including alcohol consumption, sun exposure, tobacco use, lesion, type of biopsy, lesion surface and lesion color were also collected. The dataset has a separate patches folder consisting of 3736 patches of cancerous images as well.

This dataset also consists of patients’ clinical and demographic information in a CSV as well as XLS format. Representative sample images are shown in Figure 1, and a snapshot of the dataset obtained from Kaggle [36] is provided in Figure 2. Table 2 below provides the data dictionary of clinical data. The CSV file consists of a total of 237 records with 17 columns.

1.jpg

Figure 1. Sample images

2.jpg

Figure 2. Snapshot of dataset from Kaggle [36]

Table 2. Data dictionary

Name	Column Name	Description
ID	public_id	A unique ID for each sample.
Filename	path	Histopathologic image filename.
Lesion localization	localization	Lesion’s localization in body. (Tongue, Gingiva, Buccal mucosa, Floor of mouth, Lip, Palate).
Lesion larger size	larger_size	Larger size of the lesion in centimeters.
Diagnosis	diagnosis	Lesion diagnosis. (OSCC, Leukoplakia with dysplasia, Leukoplakia without dysplasia).
Dysplasia severity	dysplasia_severity	Severity of the dysplasia in the lesion. (Mild, Severe, Moderate).
Gender	gender	Patient’s gender. Male or Female (M, F).
Age group	Age_group	Patient’s age group. Group 0 for ages lesser than 40 years. Group 1 for ages between 41 and 60 years, included. Group 2 for ages greater than 60 years.
Skin color	Skin_color	Patient’s skin color. (Black, Brown, White and Others).
Tobacco use	tobacco_use	Patient’s tobacco use. (Yes, no, Former, Not Informed).
Alcohol consumption	Alcohol_consumption	Patient’s alcohol consumption. (Yes, no, Former, Not Informed).
Sun exposure	Sun_exposure	Patient’s sun exposure in hours. It is -1, if not informed.

3.2 Computer aided detection

We proposed a computer-aided detection system that can assist the doctors in the interpretation of medical images. The system uses deep convolutional layers for feature extraction of lesion images and a multi model fusion methods in order to combine the image feature with clinical and demographic data of patients. As the doctors and expertise can diagnose the patient with their meta data and lesion image, we propose that this computer-aided detection system can assist them using the combination of information.

3.3 Model description

A deep learning approach CNN which contains some specific regions starting from lower layers to the higher layers to process raw pixel values of images. There are four basic layers of CNN: A convolutional layer for feature extraction and generating a feature map from the image provided as input, a max pooling layer that only targets the max features or pixel values and reduces the complexity, a flatten layer that converts the 2D or 3D matrices into 1D array and the dense layer which is the fully connected layer containing neurons inter connected with each other. The convolutional layer uses an activation function which adds non-linearity to the model, one of the activation functions used in our model was ReLU (Rectified Linear Unit). Using the combination of the above layers, a CNN model was created and it achieved better detection capability by tuning the hyperparameters. All the layers were available in the Keras library which were directly imported and used for image classification. Our work specifically considers a custom CNN model designed for specific multi model classification task. It is a traditional CNN architecture designed to address the unique requirements to integrate both image and clinical data for better results.

The convolutional layer applies a 3 × 3 filter on the input image of 224 × 224 to generate a feature map. The max pooling layer applies 2 × 2 matrix. Below are the steps that illustrate the formulation of the model:

• Initially, the image dataset was loaded from the directory, resized to pixel values 224 × 224, normalized the pixel values to [0,1] and converted them to an array.
• Clinical data which is the CSV file was loaded and stored into a Pandas data Frame and pre-processing was done which includes determining the categorical features, scaling numerical features to standardize them and saving the pre-processed pipeline for future use.
• The target variable was identified (Diagnosis column from the CSV file) and encoded using Label Encoder that converts them to numerical form, also the labels were one-hot encoded.
• Data was split into training and testing i.e., 80% was used for training and remaining 20% was used for testing.
• Data were trained by passing them through the layers in the form of input and as a result feature map was generated.
• ReLU activation function was used to introduce non-linearity to the model.
• After obtaining the feature map, the multimodal fusion model was generated which took two inputs: clinical data and image data. The image feature was extracted using the CNN layers and clinical data were passed through the dense layers. Both the image and clinical data were concatenated and passed through an additional number of layers in order to make the final predictions.
• The multimodal was compiled and trained using categorical cross entropy as the loss function and Adam as an optimizer. The model was trained with 200 epochs with a batch size of 32.
• The trained model was saved and a classification report was prepared along with confusion matrix and accuracy-loss graph.
• A default learning rate of 0.001 was used. To further reduce overfitting, especially since our dataset was small, we added a 50% dropout rate and L2 regularization (with λ = 0.001) to the fully connected layers.

To prepare the clinical data for training, a clear and consistent preprocessing strategy was applied. Categorical fields such as gender, tobacco use, and skin color sometimes included values like "Not Informed", instead of removing or filling these entries, they were kept and treated as valid categories to avoid losing potentially useful information. These categorical variables were then converted into numerical format using one-hot encoding, making them suitable for model input. At the same time, numerical features like lesion size and patient IDs were scaled using standard normalization so that all the values are on a similar scale. These steps were organized into a single preprocessing pipeline, which was saved for later use to ensure that the same process was carried out during testing and prediction. This helped to maintain consistency and improve the overall quality of the input data used in the multimodal model.

3.4 Model formulation

The motivation behind our research was to use a multi-model approach where histopathological images were processed by a custom CNN architecture for feature extraction, a feature map was generated at the end of the convolutional layers and was combined with clinical data. Since it’s a multimodal so the clinical data was passed at the final dense layer of the CNN architecture and combined with the feature map which was generated earlier, resulting in a merged data (i.e., both image and clinical data) for the model. Hence, the model was trained with multiple data. This approach was intended to overcome some of the challenges by being computationally effective, cost effective and providing faster training for the dataset, resulting in highly efficient diagnosis. It basically has two components. Initially the histopathological images were processed by converting the pixel values to [0,1] and were passed as input in the input layer of custom CNN layers for generation of feature map. The extracted feature map was combined with the clinical data of same patients simultaneously at the final dense layer for efficient training.

Figure 3 represents the architecture of the multimodal which was initially trained with the histopathological image dataset and the feature map was generated while passing through the custom-CNN convolutional, max pooling and flatten layers. After the feature map was successfully generated, the clinical and demographic data were introduced as input from the directory and merged with the extracted feature. The data was further trained by passing them through the dense layers for efficient result. After the model was trained efficiently an overall classification report was obtained for performance evaluation.

3.4.1 Image and clinical data fusion

The image features were combined with clinical data using a concatenate function. Below are the steps followed to concatenate both features.

Step 1: Input Data

(i) Images (224 × 224 × 3) – processed through CNN.

(ii) Clinical Data (structured/tabular format) – Processed through Dense Layer. The detailed structure of the input data used for model development is presented in Table 3.

Table 3. Input data

Sample	Image (224 × 224 × 3)	Clinical Data (Gender, Age, Smoking, etc.)
1	000.png	Male,40, Smoker, etc.
2	001.png	Female,45, Non-Smoker, etc.

Step 2: Feature Extraction

(i) CNN extracts feature from image – Output 1D vector.

(ii) Dense Layer process clinical data – Output another 1D vector. The set of extracted features derived from the dataset is summarized in Table 4.

Table 4. Extracted features

Sample	Image Features (128D)	Clinical Data (128D)
1	[0.5, 0.2, 0.7, ...]	[0.1, 0.4, 0.9, ...]
2	[0.3, 0.8, 0.6, ...]	[0.2, 0.3, 0.5, ...]

Step 3: Concatenation

(i) Both 128D vectors are merged into a single 256D feature vector.

(ii) Therefore, Final feature = Concatenate (image features, Clinical features). The concatenated representation of the multimodal data is provided in Table 5.

Figure 4 shows a custom convolutional neural network (CNN) architecture which is developed to extract meaningful features from histopathological images, each resized to 224 × 224 pixels with three color channels (RGB). The architecture, implemented in TensorFlow/Keras, is structured as follows:

1. Input Layer

This layer accepts raw images and resizes them to (224, 224, 3).

Table 5. Concatenated

Sample	Concatenated Features (256D)
1	[0.5, 0.2, 0.7, ..., 0.1, 0.4, 0.9, ...]
2	[0.3, 0.8, 0.6, ..., 0.2, 0.3, 0.5, ...]

3.jpg

Figure 3. Overall model formulation

4.jpg

Figure 4. Architecture diagram

2. Feature Extraction Blocks

The model includes three convolutional blocks, each followed by max pooling to progressively reduce spatial dimensions:

a. Block 1: 32 filters, 3 × 3 kernel, ReLU activation - MaxPooling2D (2 × 2)

b. Block 2: 64 filters, 3 × 3 kernel, ReLU activation - MaxPooling2D (2 × 2)

c. Block 3: 128 filters, 3 × 3 kernel, ReLU activation - MaxPooling2D (2 × 2)

3. Flattening Layer

The output from the final pooling layer is flattened into a 1D vector.

4. Dense Layers with Regularization

This layer captures higher-level representations and reduce overfitting, the network includes several fully connected layers with ReLU activations, dropout (0.5), and L2 regularization (λ = 0.001):

a. Dense(256) → Dropout(0.5)

b. Dense(128) → Dropout(0.5)

c. Dense(64) → Dropout(0.5)

d. Final output: Dense(128), producing a compact feature vector.

A detailed summary of the proposed model architecture is presented in Table 6.

Table 6. Model summary

Oral Cancer is a spreading medical condition which require an early detection and treatments are crucial for better outcomes. Diagnosis supported by biopsy which involves microscopic images are quiet a common approach for confirming the presence of cancer. This proposed research aims to develop a DL based mechanism for detection and classification of three types of oral cancer i.e., OSCC, leukoplakia with dysplasia and leukoplakia without dysplasia from the histopathological images and patient’s data. The main objective was to design a computer aided diagnosis system for early prediction of cancer. To achieve this, the research employed a custom designed deep learning CNN model and this model was again customized by tuning the hyperparameters such as improving the layers, epochs, target variable, batch size, etc. This experiment aims for a multimodal approach to improve the efficacy, it was termed as multimodal as both the image data and clinical data were combined so that the model gets trained collectively. Furthermore, our proposed work explored the advantages of combining the models and provided an impressive outcome with the training accuracy of ~97% and test accuracy of ~93%. The model was successful in classifying cancer from the images, indicating the value and importance of DL approaches for oral cancer diagnosis. The proposed model provides an early detection of oral cancer resulting in early treatments. This study also aims to assist the doctors and specialists by increasing the accuracy of diagnosis, reducing workload and being cost-effective.

While the model achieved a high training accuracy of 97% and a slightly lower testing accuracy of 93%, the gap suggests a mild level of overfitting. This indicates that the model may have learned some patterns which are too specific to the training data. To help reduce this effect, techniques such as dropout layers and L2 regularization were already applied. However, further improvements could include increasing the dropout rate, pre-processing techniques, adding data augmentation, or using techniques like early stopping to prevent the model from overtraining. Our future work will try to improve the model’s generalization towards testing as it will incorporate stain normalization as pre-processing technique. Data augmentation is a bit challenging as the dataset is both image as well as clinical data and both are correlated with each other. Random augmentation may lead to shuffling of data which might affect the correlation among the data.

Some of the challenges faced during this research includes data selection as it aimed for a multimodal approach which required both image data as well as clinical data. Several datasets were selected initially for the research but they were not publicly available. Secondly there were several factors which were not aligning during the merging phase. At the end we were able to coin some of the mistakes and got our research done with impressive results.

We recognize the dataset's limited size (237 samples) and uneven class distribution where there are more OSCC cases compared to leukoplakia without dysplasia. This becomes a challenge and to reduce potential bias during training, we will be using data augmentation techniques such as image rotation, flipping, and shifting to enhance sample variety and partially address class imbalance in our future work. We did not apply traditional oversampling methods, as these can introduce noise or lead to overfitting in image-based tasks. Instead, we relied on the strength and ability of modern deep learning models to generalize well, even with limited and imbalanced datasets. In future studies, we plan to explore more targeted balancing strategies or the use of synthetic data to further improve model robustness. In context with the current research carried out, it was observed that there are very few researches conducted which utilizes the importance of clinical data with image data. With this work we will try to introduce the importance of multimodal approach over single model approach and due to which the dataset used in this research is considered as most important rather the number of entities are limited to 237 only.

After considering how our multimodal model performs in testing, there are some real-world challenges to consider. The model’s accuracy may drop if the input images are of low quality, things like poor lighting, blurriness, or different camera angles can affect results. If clinical data is missing or incomplete, it may lead to less reliable predictions since the model uses both image and patient information. Differences in how data is collected across hospitals or clinics can also impact how well the model works with new settings. To overcome these challenges, it’s very important to have consistent data collection methods, strong preprocessing strategies and fine-tune the model with local data before using it in a clinical environment.

This research can be continued in the future with new approaches including a complete comparison report on various supervised and deep learning algorithms using the same dataset and more. A collective data consisting of both images as well as clinical data can be collected and various deep learning models will be trained for a collective analysis on the performance of the models. Furthermore, we aim to continue our research utilizing the technological advancements and design a system with the most efficient deep learning models.