Arsitektur Efficientnet Berbasis Transfer Learning Untuk Klasifikasi Penyakit Paru Pada Citra Chest X-Ray

Nurfaudzan, Nurfaudzan (2026) Arsitektur Efficientnet Berbasis Transfer Learning Untuk Klasifikasi Penyakit Paru Pada Citra Chest X-Ray. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6003241002-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (10MB)

Abstract

Penyakit paru masih menjadi penyebab utama morbiditas dan mortalitas global, sehingga deteksi dini melalui citra Chest X-Ray (CXR) menjadi sangat penting. Namun, interpretasi manual CXR rentan terhadap subjektivitas dan kelelahan visual, serta terbatasnya jumlah data teranotasi dan ketidakseimbangan kelas menjadi tantangan dalam pengembangan model deep learning. Penelitian ini mengimplementasikan arsitektur EfficientNet berbasis transfer learning untuk klasifikasi multi-kelas penyakit paru pada citra CXR ke dalam empat kategori: Normal, COVID-19, Pneumonia, dan Tuberkulosis. EfficientNet dipilih karena keunggulan compound scaling yang menyeimbangkan kedalaman, lebar, dan resolusi secara simultan, serta efisiensi parameternya. Tiga skenario model dikembangkan dan dibandingkan: EfficientNet-Scratch (dilatih dari awal), EfficientNet-Feature Extractor (base model dibekukan), dan EfficientNet-Fine Tuning (dua fase pelatihan). Hasil evaluasi menggunakan Macro F1-Score menunjukkan bahwa model Fine-Tuning mencapai kinerja terbaik dengan akurasi 96,16% dan Macro F1-Score 0,9603, mengungguli model Scratch (94,89%; 0,9303) dan Feature Extractor (93,89%; 0,9367). Model Fine-Tuning juga menunjukkan stabilitas pelatihan terbaik tanpa indikasi overfitting, serta berhasil mengatasi kesenjangan domain (domain shift) antara citra natural ImageNet dan citra CXR melalui strategi fine-tuning dua fase. Dari segi efisiensi komputasi, model Feature Extractor menjadi yang tercepat (131 detik), sementara Fine-Tuning menawarkan keseimbangan terbaik antara akurasi dan waktu pelatihan (354 detik). Penelitian ini membuktikan bahwa arsitektur EfficientNet dengan transfer learning mampu menghasilkan model klasifikasi penyakit paru yang akurat dan efisien secara komputasi, sehingga berpotensi diimplementasikan pada fasilitas kesehatan dengan sumber daya terbatas untuk mendukung diagnosis dini dan skrining penyakit paru.
==================================================================================================================================
Lung diseases remain a leading cause of global morbidity and mortality, making early detection through Chest X-Ray (CXR) imaging critically important. However, manual CXR interpretation is prone to subjectivity and visual fatigue, while limited annotated data and class imbalance pose challenges in deep learning model development. This study implements an EfficientNet architecture with transfer learning for multi-class classification of lung diseases on CXR images into four categories: Normal, COVID-19, Pneumonia, and Tuberculosis. EfficientNet was selected for its compound scaling advantages that simultaneously balance depth, width, and resolution, along with its parameter efficiency. Three model scenarios were developed and compared: EfficientNet-Scratch (trained from scratch), EfficientNet-Feature Extractor (base model frozen), and EfficientNet-Fine Tuning (two-phase training). Evaluation results using Macro F1-Score showed that the Fine-Tuning model achieved the best performance with 96.16% accuracy and 0.9603 Macro F1-Score, outperforming the Scratch model (94.89%; 0.9303) and Feature Extractor (93.89%; 0.9367). The Fine-Tuning model also demonstrated the best training stability without overfitting indications, successfully addressing domain shift between natural ImageNet images and CXR images through a two-phase fine-tuning strategy. In terms of computational efficiency, the Feature Extractor was the fastest (131 seconds), while Fine-Tuning offered the best balance between accuracy and training time (354 seconds). This study proves that EfficientNet architecture with transfer learning can produce accurate and computationally efficient lung disease classification models, potentially implementable in healthcare facilities with limited resources to support early diagnosis and lung disease screening.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Chest X-Ray (CXR), Deep Learning, EfficientNet, Transfer Learning, Klasifikasi Chest X-Ray (CXR), Deep learning, EfficientNet, Transfer learning, Classification
Subjects:	Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Statistics > 49101-(S2) Master Thesis
Depositing User:	Nurfaudzan Nurfaudzan
Date Deposited:	05 Aug 2026 06:04
Last Modified:	05 Aug 2026 06:04
URI:	http://repository.its.ac.id/id/eprint/143957

Available Versions of this Item

• Arsitektur Efficientnet Berbasis Transfer Learning Untuk Klasifikasi Penyakit Paru Pada Citra Chest X-Ray. (deposited 05 Aug 2026 04:37)
  • Arsitektur Efficientnet Berbasis Transfer Learning Untuk Klasifikasi Penyakit Paru Pada Citra Chest X-Ray. (deposited 05 Aug 2026 06:04) [Currently Displayed]

Actions (login required)

View Item