Electronic Theses and Dissertation

Melanoma merupakan salah satu jenis kanker kulit yang berpotensi menyebabkan kematian apabila tidak terdeteksi sejak dini. pemanfaatan deep learning pada citra dermoskopi telah menunjukkan potensi yang besar dalam membantu proses klasifikasi melanoma. namun, keberadaan artefak rambut pada citra dermoskopi dapat mengganggu proses ekstraksi fitur dan memengaruhi kinerja model. penelitian ini bertujuan untuk mengevaluasi dampak penerapan digital hair removal (dhr) terhadap kinerja dua arsitektur lightweight convolutional neural network (cnn), yaitu efficientnet-b0 dan mobilenetv2. dataset yang digunakan merupakan gabungan isic 2019 dan isic 2020 yang terdiri dari 23.114 citra benign dan malignant. proses dhr dilakukan menggunakan metode morphological black-hat dan inpainting telea. empat skenario model dibangun, yaitu efficientnet-b0 original, efficientnet-b0 hair removed, mobilenetv2 original, dan mobilenetv2 hair removed. seluruh model dilatih menggunakan transfer learning dengan bobot pra-latih imagenet. hasil penelitian menunjukkan bahwa mobilenetv2 original memberikan performa terbaik dengan accuracy 87,02%, precision 0,6524, recall 0,8796, f1-score 0,7492, dan auc-roc 0,9433. penerapan dhr memberikan dampak yang berbeda pada setiap arsitektur. pada efficientnet-b0, accuracy meningkat dari 83,28% menjadi 84,08%, sedangkan pada mobilenetv2 accuracy menurun dari 87,02% menjadi 59,98%. hasil ini menunjukkan bahwa efektivitas dhr bergantung pada karakteristik arsitektur cnn yang digunakan. sebagai implementasi, prototipe aplikasi web berbasis flask berhasil dikembangkan untuk melakukan klasifikasi melanoma secara otomatis menggunakan model terbaik hasil penelitian. kata kunci: melanoma, digital hair removal, efficientnet-b0, mobilenetv2, citra dermoskopi, transfer learning.

Melanoma is one of the most aggressive types of skin cancer and can be life-threatening if not detected at an early stage. Deep learning approaches using dermoscopic images have shown significant potential in supporting melanoma classification. However, the presence of hair artifacts in dermoscopic images may interfere with feature extraction and affect model performance. This study aims to evaluate the impact of Digital Hair Removal (DHR) on the performance of two lightweight Convolutional Neural Network (CNN) architectures, namely EfficientNet-B0 and MobileNetV2. The dataset used is a combination of the ISIC 2019 and ISIC 2020 datasets, consisting of 23,114 benign and malignant dermoscopic images. The DHR process was performed using a combination of morphological black-hat transformation and Telea inpainting. Four experimental scenarios were developed: EfficientNet-B0 Original, EfficientNet-B0 Hair Removed, MobileNetV2 Original, and MobileNetV2 Hair Removed. All models were trained using transfer learning with ImageNet pretrained weights. The results show that MobileNetV2 Original achieved the best performance, with an accuracy of 87.02%, precision of 0.6524, recall of 0.8796, F1-score of 0.7492, and AUC-ROC of 0.9433. The application of DHR produced different effects on each architecture. For EfficientNet-B0, accuracy increased from 83.28% to 84.08%, whereas for MobileNetV2, accuracy decreased from 87.02% to 59.98%. These findings indicate that the effectiveness of DHR depends on the characteristics of the CNN architecture employed. As an implementation, a Flask-based web application prototype was successfully developed to perform automatic melanoma classification using the best-performing model. Keywords: Melanoma, Digital Hair Removal, EfficientNet-B0, MobileNetV2, Dermoscopic Images, Transfer Learning.