Pengembangan MobileNetV3 Berbasis Dual Attention dalam Klasifikasi Citra Dermoskopi Kanker Kulit

Priambodo, Anas Rachmadi (2025) Pengembangan MobileNetV3 Berbasis Dual Attention dalam Klasifikasi Citra Dermoskopi Kanker Kulit. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025231044-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (6MB) | Request a copy

Abstract

Deteksi dini kanker kulit sangat penting untuk meningkatkan hasil pasien, namun metode diagnostik manual saat ini masih bersifat subjektif dan memerlukan sumber daya besar. Penelitian ini mengembangkan sistem klasifikasi multi-kelas otomatis yang efisien, berbasis MobileNetV3 yang ditingkatkan dengan Convolutional Block Attention Module (CBAM), untuk membedakan secara akurat antara lesi jinak, karsinoma sel basal, melanoma, dan karsinoma sel skuamosa menggunakan citra dermoskopi.
CBAM diintegrasikan ke dalam model pre-trained MobileNetV3. Untuk mengatasi ketidakseimbangan kelas yang ekstrem pada dataset ISIC 2024 SLICE-3D, diterapkan augmentasi data—meliputi rotasi, pembalikan, dan gangguan warna. Kinerja model dievaluasi pada set uji terpisah menggunakan metrik accuracy, precision, recall, F1-score, dan partial AUC (pAUC) pada tingkat true positive rate ≥80%. Interpretabilitas dianalisis melalui visualisasi Grad-CAM dan Score-CAM.
Model MobileNetV3+CBAM dengan augmentasi dataset berhasil mengklasifikasikan citra uji dengan akurasi 98,97% dari seluruh 1.560 sampel data, melampaui performa seluruh model yang diujikan. Nilai pAUC pada skala ter-normalisasi [0,0–0,2] berhasil dicapai 0,193. Peta atensi menunjukkan fokus yang tepat pada area lesi diagnostik.
Integrasi CBAM pada MobileNetV3 meningkatkan akurasi dengan tambahan parameter minimal. Sehingga memungkinkan inferensi real-time pada perangkat mobile atau dengan lingkungan dengan keterbatasan sumber daya perangkat.
======================================================================================================== Early detection of skin cancer is essential for improving patient outcomes, yet current manual diagnostic methods remain subjective and resource intensive. This study develops an efficient automated multiclass classification system based on MobileNetV3 enhanced with Convolutional Block Attention Module (CBAM) to accurately distinguish benign lesions, basal cell carcinoma, melanoma, and squamous cell carcinoma using dermoscopic images.
We integrated CBAM into a pretrained MobileNetV3 backbone. To address severe class imbalance in the ISIC 2024 SLICE-3D dataset, targeted data augmentation—including rotations, flips, and color perturbations—was applied. Model performance was evaluated on a held-out test set using accuracy, precision, recall, F1-score, and partial AUC (pAUC) at true positive rate ≥80%. Interpretability was assessed through Grad-CAM and Score-CAM visualizations.
The MobileNetV3+CBAM model, trained with targeted dataset augmentation, correctly classified 1,560 test images with an overall accuracy of 98.97%, outperforming all other evaluated architectures. On the normalized [0.0–0.2] partial AUC scale, the model achieved a pAUC of 0.193.
Visualization of attention maps confirmed precise localization on diagnostically significant lesion regions. By integrating CBAM with MobileNetV3 and adding fewer extra parameters, this approach supports real-time inference even on mobile devices or hardware-constrained environments.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Attention Mechanism, Convolutional Block Attention Module, Data Augmentation, MobileNetV3, Skin Cancer Classification
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Anas Rachmadi Priambodo
Date Deposited:	04 Aug 2025 12:34
Last Modified:	04 Aug 2025 12:34
URI:	http://repository.its.ac.id/id/eprint/125406

Actions (login required)

View Item