Klasifikasi Citra Endoskopi Gastrointestinal Menggunakan Transfer Learning pada Model Convolutional Neural Network

Dewa, Ghazzi (2025) Klasifikasi Citra Endoskopi Gastrointestinal Menggunakan Transfer Learning pada Model Convolutional Neural Network. Other thesis, Institut Teknologi Sepuluh Nopember.

This is the latest version of this item.

[thumbnail of 5025201074-Undergraduate_Thesis.pdf] Text
5025201074-Undergraduate_Thesis.pdf
Restricted to Repository staff only until 1 April 2027.

Download (3MB) | Request a copy

Abstract

Klasifikasi citra endoskopi gastrointestinal memiliki peran penting dalam mendeteksi dan mendiagnosis berbagai kelainan pada saluran pencernaan. Penelitian ini bertujuan untuk mengembangkan model klasifikasi menggunakan pendekatan transfer learning pada arsitektur Convolutional Neural Network , dengan model pre-trained DenseNet121, InceptionV3, dan EfficientNetB0. Dataset GastroVision dan HyperKvasir digunakan dalam penelitian ini, dilengkapi dengan berbagai teknik augmentasi data, seperti rotasi, flipping, CLAHE, random brightness/contrast, gaussian blur, dan coarse dropout, untuk meningkatkan variasi data. Hasil penelitian menunjukkan bahwa DenseNet121 memberikan performa terbaik dibandingkan model pre-trained lainnya. Dengan konfigurasi fine-tuning pada layer ke-60 dan learning rate 0,00005, model ini mencapai F1-score tertinggi sebesar 92,80% dengan loss sebesar 31,67%, mengungguli InceptionV3 dan fficientNetB0. Pada tingkat kelas, Normal Esophagus memiliki F1-score tertinggi 100%, menunjukkan deteksi yang sangat baik. Normal Z-Line mencapai 85,71%, sementara Esophagitis memiliki nilai 87,40% , yang lebih rendah akibat beberapa kesalahan klasifikasi antar dua kelas tersebut. Penelitian ini menegaskan bahwa transfer learning merupakan solusi yang efektif dalam meningkatkan akurasi klasifikasi pada dataset yang terbatas. Selain itu, penelitian ini mengidentifikasi tantangan dalam membedakan kelas dengan fitur visual yang mirip, seperti Esophagitis dan Normal Z-Line, yang masih memerlukan pendekatan lanjutan untuk peningkatan model di masa depan.
==================================================================================================================================
Gastrointestinal endoscopy image classification plays a crucial role in detecting and diagnosing abnormalities in the digestive tract. This study aims to develop a classification model using the transfer learning approach on Convolutional Neural Network architectures, including pre-trained models DenseNet121, InceptionV3, and EfficientNetB0. The GastroVision and HyperKvasir datasets were utilized, complemented with various data augmentation techniques such as rotation, flipping, CLAHE, random brightness/contrast, gaussian blur, and coarse dropout to enhance data variability. the results demonstrate that DenseNet121 outperforms other pre-trained models. With fine-tuning at the 60th layer and a learning rate of 0.00005, the model achieved the highest F1 score of 92.80% with a loss of 31.67%, surpassing nceptionV3 and EfficientNetB0. At the class level, Normal Esophagus achieved the highest F1-score 100%, indicating excellent detection. Normal Z-Line obtained 85.71%, while Esophagitis recorded 87.40%, with lower performance due to disclassifications between these two classes. This study confirms that transfer learning is an effective solution for improving classification accuracy on limited datasets. Additionally, it identifies challenges in distinguishing visually similar classes, such as Esophagitis and Normal Z-Line, which require further model enhancements for future improvements.

Item Type: Thesis (Other)
Uncontrolled Keywords: Transfer Learning, Convolutional Neural Network, Gastrointestinal, DenseNet121, Augmentasi Data., Transfer Learning, Convolutional Neural Network, Gastrointestinal, DenseNet121, Data Augmentation.
Subjects: T Technology > T Technology (General) > T57.5 Data Processing
Divisions: Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User: Ghazzi Buana Dewa
Date Deposited: 07 Feb 2025 07:09
Last Modified: 07 Feb 2025 07:09
URI: http://repository.its.ac.id/id/eprint/118558

Available Versions of this Item

Actions (login required)

View Item View Item