Generative Oversampling Berbasis Latent Blending dan Diffusion Refinement untuk Klasifikasi Citra Medis Tidak Seimbang

Santoso, Lucky (2026) Generative Oversampling Berbasis Latent Blending dan Diffusion Refinement untuk Klasifikasi Citra Medis Tidak Seimbang. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221050-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (13MB) | Request a copy

Abstract

Ketidakseimbangan kelas pada klasifikasi citra medis menyebabkan kelas minoritas yang memuat informasi diagnostik vital cenderung terabaikan. Pendekatan oversampling pada ruang laten digunakan untuk mengatasi hal ini, namun Linear Interpolation (LERP) pada metode konvensional memotong wilayah densitas rendah di interior hypersphere laten sehingga sampel sintetis menyimpang dari manifold dan mengalami penyusutan norma (norm collapse). Kelemahan ini menurunkan kualitas sampel sintetis sekaligus separabilitas antar kelas sehingga model klasifikasi tetap bias terhadap kelas mayoritas. Penelitian ini mengusulkan Generative Oversampling yang memadukan Latent Blending memakai Spherical Linear Interpolation (SLERP) untuk mempertahankan norma vektor pada permukaan hypersphere dengan Diffusion Refinement memakai Latent Diffusion Model sebagai operator proyeksi yang menarik sampel kembali ke manifold data nyata, di atas encoder VAE pretrained yang dibekukan. Sistem bersifat dual-space, yaitu laten sintetis untuk Model Tradisional dan citra hasil dekode untuk Model Visi, dengan konfigurasi final berupa SLERP, pemilihan parent berbasis klasterisasi, dan Diffusion Refinement. Evaluasi dilakukan pada tiga dataset MedMNIST v2, yaitu DermaMNIST, OCTMNIST, dan RetinaMNIST, memakai recall sebagai metrik utama serta Precision dan Recall generatif untuk menilai kualitas citra sintetis. Dibandingkan kondisi baseline tanpa oversampling, pada DermaMNIST sebagai dataset paling timpang metode usulan meningkatkan separabilitas laten kelas minoritas dan menaikkan recall Model Tradisional paling besar pada model berbasis margin dan ensemble, yaitu Random Forest dari 0,167 menjadi 0,392 dan SVM dari 0,226 menjadi 0,441, dengan peningkatan yang konsisten pada ketiga dataset. Recall Model Visi tertinggi dicapai MobileNetV3 dari 0,701 menjadi 0,757, diikuti ViT-B/16 dari 0,673 menjadi 0,730. Diffusion Refinement menaikkan kualitas citra sintetis secara serentak tanpa pertukaran antara fidelitas dan keragaman, yaitu Precision generatif dari 0,859 menjadi 0,906 dan Recall generatif dari 0,608 menjadi 0,688.
==================================================================================================================================
Class imbalance in medical image classification causes the minority classes, which carry vital diagnostic information, to be neglected. Latent space oversampling is employed to address this issue, however Linear Interpolation (LERP) in conventional approaches cuts through low-density regions in the interior of the latent hypersphere, causing synthetic samples to deviate from the manifold and undergo norm collapse. This weakness degrades both the quality of the synthetic samples and the inter-class separability, so the classifier remains biased toward the majority class. This study proposes Generative Oversampling that combines Latent Blending using Spherical Linear Interpolation (SLERP), which preserves the vector norm on the hypersphere surface, with Diffusion Refinement using a Latent Diffusion Model as a projection operator that pulls samples back to the real data manifold, on top of a frozen pretrained VAE encoder. The system is dual-space, namely synthetic latents for Traditional Models and decoded images for Vision Models, with a final configuration of SLERP, clustering-based parent selection, and Diffusion Refinement. Evaluation was conducted on three MedMNIST v2 datasets, namely DermaMNIST, OCTMNIST, and RetinaMNIST, using recall as the main metric and generative Precision and Recall to assess synthetic image quality. Compared with the baseline without oversampling, on DermaMNIST as the most imbalanced dataset, the proposed method improves the separability of minority-class latents and raises the recall of Traditional Models most for margin-based and ensemble models, with Random Forest rising from 0.167 to 0.392 and SVM from 0.226 to 0.441, with consistent gains across all three datasets. The highest Vision Model recall is achieved by MobileNetV3, rising from 0.701 to 0.757, followed by ViT-B/16 from 0.673 to 0.730. Diffusion Refinement improves synthetic image quality simultaneously without a trade-off between fidelity and diversity, with generative Precision rising from 0.859 to 0.906 and generative Recall from 0.608 to 0.688.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Generative Oversampling, Latent Blending, Diffusion Refinement, Ketidakseimbangan Data, Klasifikasi Citra Medis, Data Imbalance, Medical Image Classification.
Subjects:	T Technology > T Technology (General) > T57.5 Data Processing T Technology > T Technology (General) > T57.8 Nonlinear programming. Support vector machine. Wavelets. Hidden Markov models.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Lucky Santoso
Date Deposited:	20 Jul 2026 07:37
Last Modified:	20 Jul 2026 07:37
URI:	http://repository.its.ac.id/id/eprint/135741

Actions (login required)

View Item