 Lightweight Generative Model for Synthetic Biomedical Images with Enhanced QualityT. Taruneshwaran (),
Sanjay Chidambaram (),
Parthvi Manoj (),
S. Sakthi Swaroopan (),
Sasidharan Divya (),
V. Sowmya () and
Vinayakumar Ravi ()
A chapter in Machine Learning and Deep Learning Modeling and Algorithms with Applications in Medical and Health Care, 2025, pp 57-81 from Springer Abstract: Abstract Annotated medical imaging data, the high quality of which is often a hard-earned asset represents a big challenge in the advancement of machine learning algorithms for diagnostic applications. (rephrasing) The shortage of this (high-quality) data is strongest, above all over and above the many other limitations, when it comes to the detection and analysis of kidney cancer using CT scans, because the labeled datasets, which are possible, are not available or not enough at all due to privacy concerns, high acquisition costs, and the need for expert annotations. With the limited types of medical images being extracted from the real-world and augmented with medical and clinical data, overfitting and poor generalization of such medical AI models are evident and they are not practicable in real-world clinical settings. Therefore, in the current study, we propose to develop a new method by which the Generative Adversarial Networks (GAN) will be used for the development of synthetic kidney tumors that will be efficient in terms of the use of computational resources and will need minimum device memory and have the potential to produce images that are of a high quality. This collaborative work results in a low-complexity training platform and reduces the necessity of heavy-duty hardware, thus enabling the usage of artificial data in small-scale health care facilities and low-budget research institutions. The suggested Lightweight GAN is a fine blend of realism and efficiency and it makes use of a lightweight architecture and preprocessing techniques such as augmentation to realize the generation of lifelike, utterly dissimilar, and diagnostically useful synthetic images. Irrespective of such achievements in the new generator, the comparative image quality metrics, such as the Structural Similarity Index (SSIM) at 0.59 and the Fréchet Inception Distance (FID) at 56.05, demonstrate that the fidelity of synthetic images has been thoroughly guaranteed. In that case the Lightweight GAN is a really an impressive device which only needs a drastic reduction of 16.14 Gigs Floating-Point Operations Per Second (GFLOPs) and of 24.06 million parameters. This is achieved by cutting the computational load thus becoming a highly efficient solution which is able to be deployed in environments with limited GPU capabilities in addition the need to perform time-consuming tasks (Augmentation, formation, scaling, etc.) was avoided. Where information technology has improved, at the same time, data and artificial intelligence have entered the biology and health area. Generating the kidneys’ tumor imaging in flexible and reality conditions can diversify the data sources and hence the deep learning models become more stable and precise in the cancer detection, classification, and segmentation. Bridging this gap becomes quite more stimulating, when it comes to training cycles of AI designed for very narrow cases, where the imaging data for teaching application is too scarce and it is of too low variety and quality to be used as fake or real cases. In addition, through the inclusion of ligaments, the KIMIA interface can be integrated with medical platform simulations for radiographers and medical students allowing them to simulate on synthetic yet clinically correct datasets. Besides, Lightweight GAN’s design can be customized to work beyond the limits of kidney cancer detection, and simultaneously be used in addressing similar problems in other medical imaging techniques, such as MRI, ultrasound, and histopathology, places where the corresponding data shortage is a major challenge. The model’s accuracy provides real-time applications in AI-based diagnostic systems where physicians are assisted in deciding which option to take and better yet, in the early detection of diseases. This thesis is not only a landmark, as it paves the way to conjunge high-quality diagnostic medical image synthesis and computational feasibility, making AI-supported diagnostic solutions that are not only easily scalable but also equally or even more effective in the various clinical environments. Keywords: LightWeight Generative Adversarial Network; Synthetic data generation; Kidney tumor Computed Tomography imaging; Medical image synthesis; GAN efficiency; Computed Tomography; Structural Similarity Index; Fréchet inception distance; Resource-constrained medical applications (search for similar items in EconPapers) There are no downloads for this item, see the EconPapers FAQ for hints about obtaining it. Related works: Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text Persistent link: https://EconPapers.repec.org/RePEc:spr:ssrchp:978-3-031-98728-1_4 Ordering information: This item can be ordered from DOI: 10.1007/978-3-031-98728-1_4 Access Statistics for this chapter More chapters in Springer Series in Reliability Engineering from Springer |
Is your work missing from RePEc? Here is how to contribute.
Questions or problems? Check the EconPapers FAQ or send mail to .
EconPapers is hosted by the
School of Business at Örebro University.