Multimodal Generative Reconstruction of Oral Cancer Lesions From Sparse 2D Data Using Attention-based Deep Learning

Authors

  • Abhilash Shukla CHARUSAT
  • Dhatri Raval Faculty of Computer Science and Applications (FCA), Smt. Chandaben Mohanbhai Patel Institute of Computer Applications (CMPICA), Charotar University of Science and Technology (CHARUSAT), Changa, Gujarat, India.
  • Kalpit Soni Faculty of Computer Science and Applications (FCA), Smt. Chandaben Mohanbhai Patel Institute of Computer Applications (CMPICA), Charotar University of Science and Technology (CHARUSAT), Changa, Gujarat, India.
  • Atul Patel Faculty of Computer Science and Applications (FCA), Smt. Chandaben Mohanbhai Patel Institute of Computer Applications (CMPICA), Charotar University of Science and Technology (CHARUSAT), Changa, Gujarat, India.

DOI:

https://doi.org/10.58414/SCIENTIFICTEMPER.2026.17.6.02

Keywords:

Oral Cancer,, Multimodal Medical Imaging,, Deep Learning,, Attention-Based Fusion, Generative Reconstruction

Abstract

Early and proper diagnosis of oral cancer is essential in improving patient survival and giving an opportunity to make clinical interventions. However, the acquisition of good quality three-dimensional (3D) diagnostic information remains a major issue due to its inaccessibility, exorbitant cost, and the requirement of special imaging equipment. Clinicians are therefore inclined to use limited two-dimensional (2D) information including intraoral images and histopathological slide images that might not capture the intricate structural features of the malignant lesions. To overcome these shortcomings, the current paper suggests a new template of the generative reconstruction of oral lesions associated with cancer based on sparse and multimodal 2D images. The proposed model is based on the strength of convolutional neural networks to find strong features along with an attention-based fusion mechanism to successfully incorporate the heterogeneous data sources. The integration enables the generation of a uniform and enhanced image of the lesions without any loss of critical morphological information which is necessary to come up with an adequate diagnosis. The model is capable of learning complex spatial structures, inter-modal interactions and re-assembling very fine lesion shapes with very low data requirements. The quality of reconstruction utilizing the proposed method has been demonstrated to be much higher and affect the task of downstream cancer classification. The Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR) quantitative analysis is also successful in demonstrating that the framework is even superior to the traditional reconstruction techniques. This research broadly provides a high-scaling, low-cost tool to improve the accuracy of the diagnosis in the environment with scarce resources. The suggested framework is a hopeful outlook towards the early identification and proper treatment of oral cancer by removing the necessity to use sophisticated imaging modalities, yet securing clinical credibility.

Downloads

Download data is not yet available.

Downloads

Published

25-06-2026

Issue

Section

Research article

How to Cite

Multimodal Generative Reconstruction of Oral Cancer Lesions From Sparse 2D Data Using Attention-based Deep Learning. (2026). The Scientific Temper, 17(06), 6327-6335. https://doi.org/10.58414/SCIENTIFICTEMPER.2026.17.6.02

Similar Articles

1-10 of 587

You may also start an advanced similarity search for this article.