Klasifikasi Tingkat Keparahan Retinopati Diabetik Menggunakan Multi-Teacher Knowledge Distillation dan Ordinal Loss Berbasis CNN Ringan

Izzatina, Almira Raisa (2026) Klasifikasi Tingkat Keparahan Retinopati Diabetik Menggunakan Multi-Teacher Knowledge Distillation dan Ordinal Loss Berbasis CNN Ringan. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221250-Undergraduate_Theses.pdf - Accepted Version Restricted to Repository staff only Download (6MB) | Request a copy

Abstract

Retinopati diabetik (DR) merupakan salah satu penyebab utama kebutaan yang dapat dicegah. Meskipun deep learning efektif untuk skrining otomatis DR, model dengan akurasi tinggi umumnya memiliki kompleksitas komputasi yang besar, sedangkan model ringan memiliki keterbatasan kapasitas representasi dan sering mengabaikan sifat ordinal tingkat keparahan penyakit.
Penelitian ini mengusulkan metode Multi-Teacher Knowledge Distillation dengan hybrid loss untuk mentransfer pengetahuan dari teacher ResNet-50 dan EfficientNet-B3 ke student MobileNetV3-Small menggunakan dataset APTOS 2019 yang berisi 3.662 citra fundus retina. Hybrid loss yang diusulkan merupakan gabungan dari nominal loss berbasis cross-entropy dan ordinal loss berbasis CORAL, yang dikombinasikan dengan mengonversi output kedua teacher ke ruang probabilitas yang seragam, kemudian menggabungkannya berdasarkan tingkat kepercayaan masing-masing teacher, sehingga student dapat belajar sekaligus dari supervisi kategorik dan hubungan urutan antar tingkat keparahan penyakit.
Hasil menunjukkan bahwa multi-teacher hybrid heterogen memberikan performa terbaik dengan rata-rata akurasi 84,11%, QWK 0,8907, dan Macro F1-Score 0,7022, melampaui kedua model teacher serta multi-teacher homogen yang hanya mencapai akurasi 81,20%. Hasil ini menunjukkan bahwa kombinasi teacher yang beragam dengan hybrid loss mampu menghasilkan model klasifikasi DR yang ringan, akurat, dan sensitif terhadap urutan tingkat keparahan penyakit.
=====================================================================================================================================
Diabetic retinopathy (DR) is a leading cause of preventable blindness. While deep learning has shown strong potential for automated DR screening, high-performing models are often computationally expensive, whereas lightweight models have limited representational capacity and typically overlook the ordinal nature of DR severity levels.
This study proposes a Multi-Teacher Knowledge Distillation framework with a hybrid loss function to transfer knowledge from ResNet-50 and EfficientNet-B3 teacher models to a MobileNetV3-Small student model using the APTOS 2019 dataset, which contains 3,662 retinal fundus images. The proposed hybrid loss combines a nominal loss based on cross-entropy and an ordinal loss based on CORAL, integrated by converting the outputs of both teachers into a unified probability space and aggregating them based on each teacher's confidence level, enabling the student to simultaneously learn from categorical supervision and the ordinal relationships between DR severity levels.
The results show that the heterogeneous multi-teacher hybrid approach achieved the best performance with an accuracy of 84.11%, a Quadratic Weighted Kappa (QWK) of 0.8907, and a Macro F1-score of 0.7022, surpassing both individual teacher models and the homogeneous multi-teacher approach, which achieved an accuracy of 81.20%. These findings demonstrate that combining diverse teachers with a hybrid loss function can produce a lightweight, accurate, and ordinal-aware DR classification model.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Retinopati Diabetik, Knowledge Distillation, Multi-Teacher, Ordinal Loss, CNN Ringan.
Subjects:	T Technology > T Technology (General) T Technology > T Technology (General) > T57.8 Nonlinear programming. Support vector machine. Wavelets. Hidden Markov models.
Depositing User:	Almira Raisa Izzatina
Date Deposited:	27 Jul 2026 04:13
Last Modified:	27 Jul 2026 04:13
URI:	http://repository.its.ac.id/id/eprint/137865

Actions (login required)

View Item