Rismawan, Yudha Andriano (2026) Algoritma FLS-ACPSO untuk Optimasi Hyperparameter Support Vector Machine pada Klasifikasi Melanoma Berbasis Ekstraksi Fitur Hibrida. Masters thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 6003241044-Master_Thesis.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
6003241044-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (6MB) | Request a copy |
Abstract
Melanoma malignant adalah jenis kanker kulit yang paling mematikan dan memerlukan diagnosis dini yang akurat. Diagnosis manual pada citra dermoskopi memiliki kendala karena kemiripan antara melanoma malignant dan benign, serta adanya artefak visual dan subjektivitas pengamat. Untuk mengatasi ini, penelitian ini mengembangkan sistem Computer-Aided Diagnosis (CAD) untuk klasifikasi melanoma. Lesi diisolasi menggunakan metode Adaptive Hybrid Segmentation. Kemudian, dilakukan ekstraksi fitur klinis (Aturan ABCD) dan multi-order statistic fitur tesktur sebagai fitur handcrafted serta representasi semantik dari deep learning EfficientNet-B0. Dihasilkan fitur fusi berdimensi tinggi sehingga digunakan algoritma klasifikasi Support Vector Machine (SVM) dengan kernel Radial Basis Function (RBF) yang dioptimasi menggunakan algoritma baru, yaitu Filtration and Local Search-based Adaptive Continuous Particle Swarm Optimization (FLS-ACPSO), untuk mengestimasi hyperparameter C dan γ yang optimal. Hasil eksperimen menunjukkan bahwa metode segmentasi memperoleh kinerja lebih baik dari baseline (Otsu), baik secara kualitatif maupun kuantitatif, dengan uniformitas intra-region yang lebih tinggi dan kontras inter-region yang lebih tajam. Pada tahap ekstraksi fitur, arsitektur fusi berhasil menyintesis 1.382 fitur yang terbukti sebagai deskriptor paling diskriminatif. Dari segi optimasi, penambahan filtrasi dan strategi local search membuat FLS-ACPSO dua kali lebih efektif dan efisien dibandingkan PSO standar. Integrasi model SVM yang dioptimasi oleh FLS-ACPSO mencapai konvergensi stabil antara data latih dan uji, dengan performa pengujian yang tinggi dengan accuracy sebesar 0,8960, precision 0,9074, recall (sensitivity) 0,8820, specificity 0,9100, f1-score 0,8945, dan AUC-ROC sebesar 0,9588. Penelitian ini memberikan kontribusi berupa arsitektur CAD melanoma yang akurat, kuat, dan efisien untuk aplikasi praktis. Metode Adaptive Hybrid Segmentation meberikan hasil lebih akurat dan presisi pada citra melanoma dibandingkan metode baseline. Selain itu, algoritma FLS-ACPSO yang dikembangkan juga dapat digunakan untuk mengoptimasi permasalahan lain dalam domain kontinu.
======================================================================================================================================
Malignant melanoma is the deadliest type of skin cancer and requires precise early detection. Diagnosing malignant melanoma through dermoscopic images presents challenges due to its visual similarity to benign lesions. Additionally, visual artifacts and observer bias can influence the diagnostic outcomes. To address these issues, this study developed a Computer-Aided Diagnosis (CAD) system for melanoma classification. The initial separation of lesions was conducted utilizing the newly developed Adaptive Hybrid Segmentation method. Then, clinical features based on the ABCD Rule and multi-order statistical texture features were extracted as handcrafted features, whereas semantic representations were obtained using the EfficientNet-B0 deep learning model. The combined high-dimensional feature set was classified using a Support Vector Machine (SVM) with a Radial Basis Function (RBF) kernel. The hyperparameters (C and γ) were optimized using a new algorithm called Filtration and Local Search-based Adaptive Continuous Particle Swarm Optimization (FLS-ACPSO). Experimental results show that the proposed segmentation method is better than the baseline (Otsu) in terms of both quality and quantity, achieving higher uniformity within regions and clearer contrast between regions. In the feature extraction stage, the proposed feature fusion architecture successfully combined 1,382 features, which were the most effective descriptors. From an optimization standpoint, the addition of a filtration mechanism and a local search strategy makes FLS-ACPSO twice as effective and efficient as standard PSO. The FLS-ACPSO-optimized SVM achieved stable results between the training and testing datasets, showing excellent testing performance with an accuracy of 0.8960, precision of 0.9074, recall (sensitivity) of 0.8820, specificity of 0.9100, F1-score of 0.8945, and AUC-ROC of 0.9588. This study offers an accurate, reliable, and efficient CAD system for melanoma diagnosis. The proposed Adaptive Hybrid Segmentation method provides more accurate and precise lesion segmentation than the baseline approach. In addition, the developed FLS-ACPSO algorithm can be used for other optimization problems in continuous domains.
Actions (login required)
 |
View Item |