Peningkatan Sensitivitas Deteksi Mild Diabetic Retinopathy Menggunakan Arsitektur Hybrid CNN-GNN dengan Multitask Learning

Harjanto, Salsabila Amalia (2025) Peningkatan Sensitivitas Deteksi Mild Diabetic Retinopathy Menggunakan Arsitektur Hybrid CNN-GNN dengan Multitask Learning. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5027221023-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (8MB) | Request a copy

Abstract

Penelitian ini mengembangkan arsitektur hybrid CNN–GNN dengan Multitask Learning (MTL) untuk meningkatkan sensitivitas deteksi Mild Diabetic Retinopathy (DR) yang umumnya sulit dikenali akibat mikrolesi yang sangat halus dan ketidakseimbangan data. Sistem dibangun menggunakan EfficientNet-B0 sebagai backbone CNN, representasi graph-based dalam lima varian (Grid-4, Grid-8, KNN k=6, k=8, k=12), serta dua head prediksi (multikelas DR dan biner Mild vs Non-Mild). Pipeline meliputi pengolahan data dengan Green-CLAHE dan auto-cropping, ekstraksi fitur CNN, visualisasi Grad-CAM, pembentukan graf, message passing, joint fine-tuning CNN–GNN, dan multitask supervision. Evaluasi dilakukan melalui skema in-domain pada dataset APTOS 2019 dan cross-domain pada unseen dataset Messidor-2, dengan metrik akurasi dan recall Mild. Pada pengujian in-domain, model multitask terbaik (KNN k=8) mencapai Mild recall sebesar 0.8649 dengan peningkatan sebesar 0.1676 atau 24.7% relatif dari CNN baseline, serta akurasi kompetitif 0.7993 dengan penurunan kecil sebesar 2.4%. Pada pengujian cross-domain terhadap unseen data Messidor-2, meskipun performa keseluruhan menurun dibanding in-domain, arsitektur CNN–GNN–MTL berhasil memperoleh Mild recall sebesar 0.6280 dengan peningkatan 64.4% relatif dibanding CNN baseline dan akurasi 0.5746 dengan peningkatan sebesar 1.38%. Fold terbaik mencapai recall 0.76, jauh di atas maksimum baseline sebesar 0.52. Hasil tersebut menunjukkan bahwa integrasi GNN dan multitask supervision secara signifikan meningkatkan sensitivitas Mild DR baik di domain pelatihan maupun domain baru, menjadikan arsitektur hybrid ini lebih robust dan aplikatif dalam skenario skrining DR di dunia nyata.
====================================================================================================================================
This study develops a hybrid CNN–GNN architecture with Multitask Learning (MTL) to enhance the sensitivity of detecting Mild Diabetic Retinopathy (DR), a class that is often difficult to identify due to subtle micro-lesions and severe class imbalance. The system is built using EfficientNet-B0 as the CNN backbone, graph-based representations in five variants (Grid-4, Grid-8, KNN k=6, k=8, k=12), and two prediction heads (multiclass DR and binary Mild vs Non-Mild). The pipeline includes data preprocessing with Green-CLAHE and auto-cropping, CNN feature extraction, Grad-CAM visualization, graph construction, message passing, joint fine-tuning of the CNN–GNN, and multitask supervision. Evaluation is conducted using an in-domain scheme on the APTOS 2019 dataset and a cross-domain scheme on the unseen Messidor-2 dataset, using accuracy and Mild recall as metrics. In the in-domain evaluation, the best multitask model (KNN k=8) achieves a Mild recall of 0.8649, representing an improvement of 0.1676 or 24.7% relative to the CNN baseline, along with a competitive accuracy of 0.7993, showing only a minor reduction of 2.4%. In the cross-domain evaluation on the unseen Messidor-2 dataset, although the overall performance decreases compared to the in-domain results, the CNN–GNN–MTL architecture obtains a Mild recall of 0.6280, corresponding to a 64.4% relative improvement over the CNN baseline, and an accuracy of 0.5746, representing a 1.38% improvement. The best-performing fold reaches a recall of 0.76, substantially higher than the baseline maximum of 0.52. These results demonstrate that incorporating GNNs and multitask supervision significantly enhances Mild DR sensitivity in both the training domain and new, unseen domains, making the hybrid architecture more robust and applicable for real-world DR screening scenarios.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Diabetic Retinopathy, Mild DR, Deep Learning, CNN, GNN, Multitask Learning, Grad-CAM, Diabetic Retinopathy, Mild DR, Deep Learning, CNN, GNN, Multitask Learning, Grad-CAM
Subjects:	T Technology > T Technology (General) > T385 Visualization--Technique T Technology > T Technology (General) > T57.5 Data Processing T Technology > T Technology (General) > T58.62 Decision support systems T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information Technology > 59201-(S1) Undergraduate Thesis
Depositing User:	Harjanto Salsabila Amalia
Date Deposited:	19 Jan 2026 05:31
Last Modified:	19 Jan 2026 05:31
URI:	http://repository.its.ac.id/id/eprint/129712

Actions (login required)

View Item