SARI, Ira Puspita and Warsito, Budi and Nurhayati, Oky Dwi (2026) ADAPTIVE CONFIDENCE–ENTROPY WEIGHTED ENSEMBLE CNN BERBASIS TRANSFER LEARNING UNTUK KLASIFIKASI TELUR CACING PADA CITRA MIKROSKOPIS. Doctoral thesis, UNIVERSITAS DIPONEGORO.

	Text cover aja.pdf Download (175kB)
	Text Cover lengkap.pdf Restricted to Repository staff only Download (1MB)
	Text Bab 1.pdf Download (301kB)
	Text Bab 2.pdf Download (2MB)
	Text Bab 3.pdf Restricted to Repository staff only Download (1MB)
	Text Bab 4.pdf Restricted to Repository staff only Download (3MB)
	Text Bab 5.pdf Restricted to Repository staff only Download (251kB)
	Text Daftar pustaka.pdf Download (286kB)

Abstract

Diagnosis infeksi cacing parasit umumnya dilakukan melalui pemeriksaan mikroskopis telur cacing secara manual, yang sangat bergantung pada keahlian analis laboratorium serta rentan terhadap kesalahan akibat kemiripan morfologi antar spesies (fine-grained problem) dan variasi kualitas citra. Penelitian ini bertujuan mengembangkan sistem klasifikasi otomatis telur cacing berbasis Convolutional Neural Network (CNN) dengan pendekatan transfer learning dan meningkatkan kinerjanya melalui strategi ensemble adaptif berbasis tingkat keyakinan dan ketidakpastian model. Dataset yang digunakan terdiri dari 11.000 citra mikroskopis telur cacing yang dikelompokkan ke dalam 11 kelas. Tiga arsitektur CNN pra-latih, yaitu EfficientNetB0, MobileNetV3Large, dan ResNet50, dilatih dan dievaluasi menggunakan skema fine-tuning. Kontribusi utama penelitian ini adalah pengembangan framework Adaptive Multi-Backbone Ensemble yang menggabungkan keluaran beberapa model CNN secara dinamis berdasarkan nilai confidence dan entropy pada setiap sampel. Model dengan tingkat keyakinan tinggi dan ketidakpastian rendah diberi bobot lebih besar dalam proses pengambilan keputusan. Pendekatan ini tidak memerlukan pelatihan ulang model sehingga lebih efisien untuk implementasi. Hasil pengujian pada data uji menunjukkan bahwa metode yang diusulkan meningkatkan akurasi klasifikasi yang lebih tinggi dibandingkan model tunggal maupun soft voting konvensional. Analisis statistik menggunakan uji McNemar dan interval kepercayaan menunjukkan bahwa metode adaptif memberikan performa yang stabil dengan pengurangan kesalahan pada kasus-kasus yang sulit diklasifikasikan. Penelitian ini menunjukkan bahwa pendekatan ensemble adaptif berbasis ketidakpastian efektif untuk meningkatkan robustness pada klasifikasi fine-grained citra medis dan berpotensi mendukung sistem diagnosis otomatis infeksi cacing parasit berbasis citra mikroskopis.
Kata kunci: Klasifikasi, CNN, Ensemble Adaptif, Confidence, Entropy, Transfer Learning, Telur Cacing

Diagnosis of parasitic worm infections is generally performed through manual mi- croscopic examination of worm eggs, which heavily depends on the expertise of laboratory analysts and is prone to errors due to morphological similarities between species (a fine-grained problem) and variations in image quality. This study aims to develop an automatic worm egg classification system based on Convolutional Neural Networks (CNNs) using a transfer learning approach and to enhance its performance through an adaptive ensemble strategy based on model confidence and uncertainty. The dataset consists of 11,000 microscopic images of worm eggs categorized into 11 classes. Three pre-trained CNN architectures which is EfficientNetB0, MobileNetV3Large, and ResNet50 were trained and evaluated using a fine-tuning scheme. The main contribution of this research is the development of an Adaptive Multi-Backbone Ensemble framework, which dynamically combines outputs from multiple CNN Models based on confidence and entropy for each sample. Models with higher confidence and lower uncertainty are assigned greater weight in the decision-making process. This approach does not require retraining of the models, making it efficient for practical implementation. Testing on the held-out dataset shows that the proposed method improves classification accuracy outperforming than single models and conventional soft voting. Statistical Analyses using the McNemar Test and confidence intervals indicate that the adaptive method provides stable performance with reduced errors in challenging cases. This study demonstrates that an uncertainty-based adaptive ensemble approach is effective in enhancing robustness for fine-grained medical image classification and has the potential to support automated diagnostic systems for parasitic worm infections using microscopic images.
Keywords: Classification, Convolutional Neural Network, Adaptive Ensemble, Confidence, Entropy, Transfer Learning, Worm Eggs

Actions (login required)

View Item