Perbaikan Citra Medis Low-Resolution Menggunakan REAL-ESRGAN

Hidayanto, Atha Dzaky (2026) Perbaikan Citra Medis Low-Resolution Menggunakan REAL-ESRGAN. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025021269_Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (8MB) | Request a copy

Abstract

Citra medis memegang peranan vital dalam diagnosis klinis dan penanganan pasien, namun kualitasnya sering kali terdegradasi akibat keterbatasan perangkat, noise, dan resolusi rendah yang menghambat kemudahan dokter dalam menganalisis citra. Penelitian ini bertujuan mengatasi permasalahan tersebut dengan mengembangkan model perbaikan citra menggunakan Real-Enhanced Super-Resolution Generative Adversarial Network (Real-ESRGAN) yang dioptimalkan melalui metode fine-tuning pada domain medis. Metodologi yang diterapkan meliputi strategi transfer learning dari model pre-trained pada dataset DIV2K, yang kemudian diadaptasi secara spesifik pada dataset retina STARE dan citra X-Ray paru-paru Shenzhen Tuberculosis menggunakan skema degradasi bicubic downsampling dengan faktor skala ×4 untuk menjaga integritas struktur anatomi. Hasil evaluasi kinerja menunjukkan bahwa model Real-ESRGAN Fine-Tuned secara konsisten mengungguli metode interpolasi Bicubic, ESRGAN, dan Real-ESRGAN Base, di mana pada dataset STARE model ini mencapai nilai rata-rata Peak Signal-to-Noise Ratio (PSNR) tertinggi sebesar 37.92 dB dan Structural Similarity Index Measure (SSIM) 0.9007. Sementara itu, pada dataset Shenzhen TB, model juga mencatatkan performa terbaik dengan nilai rata-rata PSNR sebesar 33.41 dB dan SSIM 0.8178. Analisis visual menunjukkan bahwa model ini mampu menghapus artefak jaring (grid artifacts) dan mengurangi halusinasi fitur yang terlihat pada model lain, serta berhasil memperjelas detail pembuluh darah dan struktur tulang secara signifikan. Kesimpulannya, metode fine-tuning pada Real-ESRGAN terbukti berhasil dalam meningkatkan kualitas resolusi citra medis sambil menjaga keaslian informasi klinis, sehingga berpotensi mendukung sistem diagnosis berbasis komputer (Computer-Aided Diagnosis).
=========================================================================================================================================
Medical images play a vital role in clinical diagnosis and patient management however, their quality is often degraded due to hardware limitations, noise, and low resolution, which hinder physicians in accurately analyzing the images. This research aims to address these issues by developing an image enhancement model using the Real-Enhanced SuperResolution Generative Adversarial Network (Real-ESRGAN), optimized through fine-tuning on the medical imaging domain. The proposed methodology employs a transfer learning strategy from a pre-trained model on the DIV2K dataset, which is subsequently adapted specifically to the STARE retinal dataset and the Shenzhen Tuberculosis chest X-ray dataset using a bicubic downsampling degradation scheme with a scaling factor of ×4 to preserve anatomical structural integrity. The performance evaluation results demonstrate that the fine-tuned RealESRGAN model consistently outperforms Bicubic interpolation, ESRGAN, and the RealESRGAN Base model, achieving the highest average Peak Signal-to-Noise Ratio (PSNR) of 37.92 dB and Structural Similarity Index Measure (SSIM) of 0.9007 on the STARE dataset. Meanwhile, on the Shenzhen TB dataset, the model also records the best performance with an average PSNR of 33.41 dB and an SSIM of 0.8178. Visual analysis indicates that the proposed model effectively removes grid artifacts and reduces feature hallucination observed in other models, while significantly enhancing the clarity of blood vessel details and bone structures. In conclusion, the fine-tuning approach applied to Real-ESRGAN is proven to improve the resolution quality of medical images while preserving clinical information authenticity, thereby potentially supporting computer-aided diagnosis (CAD) systems.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Citra Medis, Real-ESRGAN, Fine-Tuning, PSNR, SSIM.
Subjects:	Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) R Medicine > R Medicine (General) > R858 Deep Learning
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	ATHA DZAKY HIDAYANTO
Date Deposited:	30 Jan 2026 06:51
Last Modified:	30 Jan 2026 06:51
URI:	http://repository.its.ac.id/id/eprint/131299

Actions (login required)

View Item