Electronic Theses and Dissertation

ABSTRAK

Klasifikasi citra Magnetic Resonance Imaging (MRI) tumor otak memerlukan model yang tidak hanya memiliki performa tinggi pada data uji internal, tetapi juga mampu mempertahankan performa pada data dari sumber berbeda. Penelitian ini bertujuan mengevaluasi dan membandingkan model DenseNet201 dan ResNet-50 dalam mengklasifikasikan citra MRI ke dalam empat kelas, yaitu glioma, meningioma, pituitary, dan no tumor, ditinjau dari performa klasifikasi, efisiensi komputasi, dan kemampuan generalisasi. Dataset utama terdiri atas 7.200 citra. Setelah dilakukan penghapusan duplikat, tersisa 7.013 citra pada data pelatihan, validasi, dan pengujian. Kedua model dikembangkan menggunakan transfer learning dan hyperparameter tuning, kemudian dievaluasi pada data uji internal serta sepuluh dataset eksternal tanpa pelatihan ulang. Audit overlap dilakukan menggunakan Message-Digest Algorithm 5 (MD5), average hash (aHash), perceptual hash (pHash), dan difference hash (dHash) untuk membantu interpretasi hasil eksternal. DenseNet201 V18 memperoleh accuracy sebesar 96,60% dan macro F1-score sebesar 96,53%, sedangkan ResNet-50 V7 memperoleh accuracy sebesar 95,22% dan macro F1-score sebesar 95,07%. ResNet-50 V7 menghasilkan waktu inferensi Central Processing Unit (CPU) yang lebih cepat, yaitu 384,60 ms per citra dibandingkan DenseNet201 V18 sebesar 397,87 ms. Pada sepuluh dataset eksternal, rata-rata akurasi keseluruhan DenseNet201 V18 mencapai 82,79%, sedangkan ResNet-50 V7 mencapai 81,82%. Variasi hasil eksternal menunjukkan bahwa performa perlu ditafsirkan bersama perbedaan format kelas, modalitas citra, dan tingkat overlap. DenseNet201 V18 dipilih sebagai model utama dan diimplementasikan bersama ResNet-50 V7 sebagai model pembanding pada prototipe website sederhana yang menampilkan hasil klasifikasi dan visualisasi Explainable Artificial Intelligence (XAI).

Kata kunci: Klasifikasi tumor otak, DenseNet201, ResNet-50, Dataset eksternal, Overlap, Explainable Artificial Intelligence

ABSTRACT Brain tumor classification using Magnetic Resonance Imaging (MRI) requires models that perform well on internal test data and retain performance on data from different sources. This study evaluates and compares DenseNet201 and ResNet-50 for classifying MRI images into four classes: glioma, meningioma, pituitary, and no tumor, based on classification performance, computational efficiency, and generalization ability. The dataset comprised 7,200 images. After duplicates were removed, 7,013 images remained across the training, validation, and test sets. Both models were developed using transfer learning and hyperparameter tuning, then evaluated on the internal test set and ten external datasets without retraining. An overlap audit using Message-Digest Algorithm 5 (MD5), average hash (aHash), perceptual hash (pHash), and difference hash (dHash) was conducted to support interpretation of the external results. DenseNet201 V18 achieved 96.60% accuracy and a 96.53% macro F1-score, whereas ResNet-50 V7 achieved 95.22% accuracy and a 95.07% macro F1-score. ResNet-50 V7 produced a faster Central Processing Unit (CPU) inference time of 384.60 ms per image, compared with 397.87 ms for DenseNet201 V18. Across the ten external datasets, DenseNet201 V18 achieved a mean overall accuracy of 82.79%, while ResNet-50 V7 achieved 81.82%. The variation in external results indicates that performance should be interpreted alongside differences in class formats, imaging modalities, and overlap levels. DenseNet201 V18 was selected as the primary model and implemented alongside ResNet-50 V7 as a comparison model on a simple website prototype that presents classification results and Explainable Artificial Intelligence (XAI) visualizations. Keywords: Brain tumor classification, DenseNet201, ResNet-50, External datasets, Overlap, Explainable Artificial Intelligence