Analisis Continual Learning Untuk Klasifikasi Citra Medis

Budianto, Timothy Hosia (2025) Analisis Continual Learning Untuk Klasifikasi Citra Medis. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025211098-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (12MB) | Request a copy

Abstract

Selama beberapa tahun terakhir, deep learning telah berhasil diterapkan di berbagai bidang termasuk klasifikasi patologi pada citra x-ray. Meski begitu keinginan untuk model memperluas area klasifikasi dengan keterbatasan akses data lamanya, keterbatasan komputasional, dan batasan kode etik pada data medis memberikan masalah baru yang disebut catastrophic forgetting. Catastrophic forgetting adalah keadaan di mana model melupakan kemampuan sebelumnya yang disebabkan karena beradaptasi dengan masuknya aliran data baru, hal tersebut sangat berbahaya apalagi pada konteks medis. Pembelajaran berkelanjutan (continual learning) dapat muncul sebagai solusi untuk tantangan ini, yang memungkinkan model untuk beradaptasi dengan data baru sambil mempertahankan pengetahuan sebelumnya. Berbagai Strategi continual learning telah dikembangkan, beberapa di antaranya yaitu pelatihan gabungan (Joint Training) sebagai baseline, metode replay (sample-based dan generative), teknik berbasis regulasi (elastic weight consolidation dan learning without forgetting), serta pendekatan arsitektur (incremental classifier). Penelitian ini mengimplementasi dan mengevaluasi Strategi tersebut pada skenario pembelajaran klasifikasi inkremental menggunakan multi-head model DenseNet-121 dengan dataset ChestX-ray14 NIH melalui framework continual learning Avalanche. Hasil evaluasi menunjukkan bahwa pendekatan berbasis memori, khususnya Experience Replay yang dioptimasi, memberikan performa terbaik dengan Accuracy 0,7452 dan forgetting rate 0,0423, memberikan peningkatan 81,1% Accuracy dan pengurangan 88,7% forgetting dibandingkan baseline Naive (0,4115 Accuracy , 0,3744 forgetting). Optimisasi hyperparameter menggunakan AdamW optimizer dengan learning rate 0,0001 terbukti memberikan peningkatan signifikan 40,6% pada combined score, dengan konfigurasi optimal mencapai gap hanya 4,6% dari theoretical upper bound (Joint Training: 0,7814 Accuracy ). Temuan ini memberikan panduan implementasi praktis dan wawasan tentang bagaimana Strategi pembelajaran berkelanjutan bekerja dalam meningkatkan model klasifikasi citra medis dengan menetapkan benchmark baru untuk penelitian selanjutnya.
=========================================================================================================================================
Over the past few years, deep learning has been successfully applied in various fields including pathology classification in x-ray images. However, the desire for models to expand the classification area with limited access to old data, computational limitations, and ethical restrictions on medical data presents a new problem called catastrophic forgetting. Catastrophic forgetting is a state where the model forgets its previous abilities due to adapting to the entry of new data streams, which is very dangerous especially in the medical context. Continual learning can emerge as a solution to this challenge, allowing the model to adapt to new data while maintaining its previous knowledge. Various continual learning Strategies have been developed, some of which are Joint Training as a baseline, replay methods (sample-based and generative), regulation-based techniques (elastic weight consolidation and learning without forgetting), and architectural approaches (incremental classifier). This study implements and evaluates these Strategies in an incremental classification learning scenario using the DenseNet-121 multi-head model with the NIH ChestX-ray14 dataset through the Avalanche continual learning framework. The evaluation results show that the memory-based approach, especially the Optimized Experience Replay, provides the best performance with an Accuracy of 0,7452 and a forgetting rate of 0,0423, providing an 81,1% increase in Accuracy and an 88,7% reduction in forgetting compared to the Naive baseline (0,4115 Accuracy , 0,3744 forgetting). Hyperparameter optimization using AdamW optimizer with Learning Rate 0,0001 is proven to provide a significant improvement of 40,6% in Combined Score, with the optimal configuration achieving a gap of only 4,6% from the theoretical upper bound (Joint Training: 0,7814 Accuracy ). These findings provide practical implementation guidelines and insights into how continual learning Strategies work in improving medical image classification models by setting a new benchmark for future research.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Pembelajaran Berkelanjutan, Kelupaan Katastrofik, Klasifikasi Citra Medis, DenseNet-121, Avalanche Framework, Elastic Weight Consolidation, Learning Without Forgetting, Replay Methods, Incremental Classifier, Continual learning, Catastrophic forgetting, Medical Image Classification
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 > T Technology (General) > T59.7 Human-machine systems.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Timothy Hosia Budianto
Date Deposited:	31 Jul 2025 06:15
Last Modified:	31 Jul 2025 06:15
URI:	http://repository.its.ac.id/id/eprint/124057

Actions (login required)

View Item