%0 Thesis
%9 Skripsi
%A Rizki Surya Nugroho, NIM.: 22106050019
%B FAKULTAS SAINS DAN TEKNOLOGI
%D 2026
%F digilib:76859
%I UIN SUNAN KALIJAGA YOGYAKARTA
%K Fusi Multimodal, Tumor Otak, Evidential Neural Network, Teori Dempster-Shafer, Contextual Correction, Interval Dominance
%P 112
%T PENGEMBANGAN MODEL FUSI CITRA MEDIS MULTIMODAL UNTUK APLIKASI DETEKSI TUMOR OTAK MENGGUNAKAN EVIDENTIAL NEURAL NETWORK DAN MEKANISME CONTEXTUAL CORRECTIONS BERBASIS BELIEF FUNCTION FRAMEWORK
%U https://digilib.uin-suka.ac.id/id/eprint/76859/
%X This study develops a multimodal medical image fusion model combining Magnetic Resonance Imaging (MRI) and Computed Tomography (CT scans) for brain tumor detection applications. Combining these two modalities aims to overcome the limitations of single-modality analysis, as MRI excels in soft-tissue contrast resolution while CT scans are more responsive to specific density structures. However, information fusion frequently faces challenges of ambiguity and conflicting information due to varying quality and reliability levels across sources. To manage this uncertainty mathematically and explicitly, this study proposes an Evidential Neural Network (ENN) approach based on the Dempster-Shafer (belief function) theoretical framework. Feature extraction was performed using a ResNet24 architecture on a balanced Kaggle dataset comprising 4,000 images per modality (consisting of 2,000 Healthy and 2,000 Tumor images). Because source reliability is not always uniform, this study implemented contextual correction mechanisms, which include contextual discounting, reinforcement, and negating. Comprehensive evaluations showed that global parameter correction was ineffective. Therefore, an adaptive strategy based on decision areas (decision area-based) formed through the interval dominance criterion was proposed. Information fusion was then conducted at the decision level using Dempster's rule of combination. Robustness testing using 10-fold cross-validation proved that this multimodal fusion approach is highly robust, drastically reducing the Euclidean Plausibility Loss (EPL) to an average of 30.2157, and improving decision quality metrics utility-discounted accuracy U65 to an average of 0.9517 and U80 to 0.9518. The model is proven capable of representing uncertainty precisely and significantly improving classification performance compared to single-modality usage.
%Z Siti Mutmainah, S.Kom, M.Cs. Ph.D.