Integrasi Mekanisme Attention pada Backbone CNN-Based Model Untuk Klasifikasi Penyakit Kanker Kolorektal Pada Citra Endoskopi

Putra, Gregorius Guntur Sunardi (2025) Integrasi Mekanisme Attention pada Backbone CNN-Based Model Untuk Klasifikasi Penyakit Kanker Kolorektal Pada Citra Endoskopi. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025231023-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (7MB) | Request a copy

Abstract

Luminal cancer terdiri dari kanker esofagus, lambung dan kolorektal menjadi salah satu penyebab kematian tertinggi di dunia akibat kanker, lebih dari dua juta jiwa pada tahun 2022. Diagnosis dini penting untuk mengurangi tingkat mortalitas kanker, dengan jumlah data yang besar proses diagnosis cenderung lebih lama dan memakan waktu. Sehingga dikembangkan sistem Computer-aided Diagnosis berbasis Convolutional Neural Networks (CNN) untuk mendiagnosis penyakit berbasis data visual. Akan tetapi CNN masih kurang fokus dalam mengenali informasi penting khususnya pada data dengan citra dengan kemiripan yang tinggi antar kelas.
Penelitian ini menganalisis integrasi mekanisme attention pada CNN dalam meningkatkan feature learning. Attention mampu memberikan titik fokus dengan memberikan bobot pada informasi relevan berdasarkan korelasi channel ataupun spasial. Mekanisme attention yang digunakan yaitu Squeeze-and-Excitation (SE), Convolutional Block Attention Module (CBAM), Efficient Channel Attention (ECA), External Attention (EA), dan Spatial Group-wise Enhance (SGE). Mekanisme attention diintegrasikan pada feature learning CNN, dengan memodifikasi arsitektur ResNet dan EfficientNet. Pengujian dilakukan pada dataset DCMH dengan 3000 gambar yang terbagi ke dalam kelas Kanker Kolorektal, Kolon Polip, dan Kolon Normal.
Hasil penelitian ini menunjukkan bahwa integrasi mekanisme attention ke dalam CNN meningkatkan kinerja hingga 9% dibandingkan tanpa mekanisme perhatian. EfficientNetB0 + SE adalah model dengan kinerja tertinggi dengan akurasi 85,8%, presisi 85,3%, MCC 0,773 diikuti oleh EfficientNetV2S + SGE dengan akurasi 85,5%, presisi 85,4%, dan MCC 0,770. Dalam memprediksi kelas kanker model EfficientNetV2S + SGE lebih baik dibandingkan model EfficentNetB0 + SE. Mekanisme attention telah menunjukkan peningkatan kemampuan feature learning pada CNN akan tetapi dalam beberapa kasus menunjukkan penurunan yang disebabkan terjadinya overfitting dan jumlah data latih yang sedikit.

============================================================================================================================

Luminal cancer, consisting of esophageal, gastric, and colorectal cancers, is one of the leading causes of cancer death worldwide, with more than two million deaths occurring in 2022. Early diagnosis is crucial for reducing cancer mortality rates, as large amounts of data make the diagnostic process more time-consuming. Therefore, a Computer-Aided Diagnosis system based on Convolutional Neural Networks (CNN) was developed to diagnose diseases based on visual data. However, CNN still lacks focus in recognizing important information, especially in data with images with high similarity between classes.
This study analyzes the integration of attention mechanisms in CNNs to improve feature learning. Attention can provide a focal point by giving weight to relevant information based on channel or spatial correlation. The attention mechanisms used are Squeeze-and-Excitation (SE), Convolutional Block Attention Module (CBAM), Efficient Channel Attention (ECA), External Attention (EA), and Spatial Group-wise Enhancement (SGE). The attention mechanism is integrated into CNN feature learning by modifying the ResNet and EfficientNet architectures. Testing was conducted on the DCMH dataset with 3,000 images divided into Colorectal Cancer, Colon Polyp, and Normal Colon classes.
The results of this study indicate that integrating the attention mechanism into a CNN improves performance by up to 9% compared to the CNN without the attention mechanism. EfficientNetB0 + SE was the highest-performing model with 85.8% accuracy, 85.3% precision, and an MCC of 0,773, followed by EfficientNetV2S + SGE with 85.5% accuracy, 85.4% precision, and an MCC of 0,770. In predicting cancer class, the EfficientNetV2S + SGE model performed better than the EfficientNetB0 + SE model. The attention mechanism has shown improvements in CNN feature learning capabilities, but in some cases, it has shown declines due to overfitting and a small amount of training data.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Attention, Biomedis, Endoskopi, Kanker Kolorektal, Saluran Pencernaan
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QR Microbiology > QR 201.T84 Tumors. Cancer R Medicine > R Medicine (General) > R858 Deep Learning T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing. T Technology > TA Engineering (General). Civil engineering (General) > TA174 Computer-aided design.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Gregorius Guntur Sunardi Putra
Date Deposited:	31 Jul 2025 06:57
Last Modified:	31 Jul 2025 06:57
URI:	http://repository.its.ac.id/id/eprint/124857

Actions (login required)

View Item