Klasifikasi Kanker Kulit Berbasis Citra dan Data Demografi Menggunakan EfficientNetV2 dan Swin Transformer

Arya Veda Alviantoro, Dr. Dyah Aruming Tyas, S.Si.

2025 | Skripsi | ELEKTRONIKA DAN INSTRUMENTASI

Klasifikasi kanker kulit merupakan tantangan penting dalam diagnosis dini, di mana deep learning menawarkan potensi solusi non-invasif. Namun, evaluasi sistematis performa arsitektur modern seperti EfficientNetV2 dan Swin Transformer, terutama dampak integrasi data demografi klinis dan pengaruh skala dataset ISIC, masih terbatas, sehingga mendorong dilakukannya penelitian ini untuk menganalisis faktor-faktor tersebut.

Penelitian ini mengevaluasi varian EfficientNetV2 (S, M, L) dan model hibrida (EfficientNetV2-S + Swin-Base) menggunakan dataset citra dermoskopi ISIC (gabungan 2017-2020 dan ISIC 2020 saja) serta data demografi klinis pasien (diolah via MLP). Model dilatih menggunakan transfer learning dengan strategi bertahap yang adaptif dan dievaluasi berdasarkan metrik seperti AUC Biner/Multikelas dan Akurasi Biner/Multikelas, dilengkapi analisis interpretabilitas Grad-CAM.

Hasil utama menunjukkan integrasi data demografi secara signifikan meningkatkan kemampuan diskriminatif (misalnya, AUC Biner hingga 0.9509 dengan EffNetV2-M+Meta) dan presisi, meski sedikit mengorbankan recall biner. Pelatihan pada dataset gabungan ISIC (~58rb citra) terbukti esensial untuk generalisasi robust, jauh mengungguli model yang dilatih hanya pada ISIC 2020 (~33rb citra). Ensemble learning heterogen (post-hoc) lebih lanjut meningkatkan kinerja, mencapai AUC Biner 0.9615 dan Akurasi Biner 95.13%.

Skin cancer classification is a critical challenge for early diagnosis, where deep learning offers potential non-invasive solutions. However, systematic evaluations of modern architectures like EfficientNetV2 and Swin Transformer, particularly regarding the impact of clinical data demografi integration and ISIC dataset scale, remain limited, motivating this research to analyze these factors.

This study evaluates EfficientNetV2 variants (S, M, L) and a hybrid model (EfficientNetV2-S + Swin-Base) using ISIC dermoscopy image datasets (combined 2017-2020 and ISIC 2020 only) along with patient clinical data demografi (processed via MLP). Models were trained using transfer learning with an adaptive staged strategy and evaluated based on key metrics such as Binary/Multiclass AUC and Accuracy, supplemented by Grad-CAM for interpretability analysis.

The main results indicate that data demografi integration significantly enhances discriminative ability (e.g., Binary AUC up to 0.9509 with EffNetV2-M+Meta) and precision, albeit at the cost of slightly reduced binary recall. Training on the larger combined ISIC dataset (~58k images) proved essential for robust generalization, vastly outperforming models trained solely on ISIC 2020 (~33k images). Post-hoc heterogeneous ensemble learning further boosted performance, achieving a Binary AUC of 0.9615 and 95.13% Binary Accuracy.

Kata Kunci : Klasifikasi Kanker Kulit, Deep Learning, EfficientNetV2, Swin Transformer, Integrasi Data Demografi, Dataset ISIC, Ensemble Learning