Please use this identifier to cite or link to this item:
https://repositori.uma.ac.id/handle/123456789/27662
Title: | Enchancing Brain Tumor Disease Classification via SqueezeNet Architecture Integrated with Group Convolution |
Other Titles: | Enchancing Brain Tumor Disease Classification via SqueezeNet Architecture Integrated with Group Convolution |
Authors: | Gultom, William |
metadata.dc.contributor.advisor: | Muhatir |
Keywords: | Brain Tumor Classification;MRI;CNN;SqueezeNet;Group Convolution;Klasifikasi Tumor Otak |
Issue Date: | Apr-2025 |
Publisher: | Universitas Medan Area |
Series/Report no.: | NPM;208160003 |
Abstract: | Klasifikasi tumor otak berbasis citra MRI merupakan tantangan signifikan dalam pengolahan citra medis, terutama ketika menghadapi ketidakseimbangan jumlah data antar kelas. Ketimpangan ini dapat menyebabkan model bias terhadap kelas mayoritas dan menurunkan sensitivitas terhadap kelas minoritas, yaitu pasien dengan tumor. Penelitian ini bertujuan untuk menganalisis pengaruh penerapan teknik Group Convolution pada arsitektur VGG19 dan SqueezeNet dalam meningkatkan efisiensi dan akurasi klasifikasi tumor otak. Penelitian menggunakan pendekatan kuantitatif eksperimental dengan implementasi Convolutional Neural Network (CNN) berbasis PyTorch. Dataset yang digunakan terdiri dari dua kelas, “Yes” (dengan tumor) dan “No” (tanpa tumor), yang dibagi ke dalam folder Train, Validation, dan Test. Model diuji dengan membandingkan performa antara arsitektur standar dan versi modifikasi yang mengintegrasikan Group Convolution. Hasil eksperimen menunjukkan bahwa SqueezeNet dengan Group Convolution berhasil mencapai akurasi hingga 90%, lebih tinggi dibandingkan model orisinal. Selain itu, sensitivitas terhadap kelas minoritas meningkat secara signifikan, menunjukkan kemampuan model dalam menangani data tidak seimbang. Temuan ini mengindikasikan bahwa Group Convolution dapat meningkatkan efisiensi komputasi sekaligus memperbaiki performa klasifikasi. Dengan demikian, teknik ini relevan untuk pengembangan sistem diagnosis otomatis. Penelitian selanjutnya disarankan menggabungkan pendekatan ini dengan metode seperti attention mechanism untuk hasil klasifikasi yang lebih optimal dan andal. Brain tumor classification using MRI images is a major challenge in medical image processing, particularly when facing imbalanced data between classes. This imbalance often leads to model bias toward the majority class and reduces sensitivity to the minority class—patients with tumors. This study aims to analyze the impact of applying Group Convolution techniques to the VGG19 and SqueezeNet architectures to enhance both computational efficiency and classification accuracy. A quantitative experimental approach was employed, implementing Convolutional Neural Networks (CNNs) using the PyTorch framework. The dataset includes two classes, “Yes” (with tumor) and “No” (without tumor), organized into Train, Validation, and Test folders. The models were evaluated by comparing the performance of standard architectures with modified versions integrating Group Convolution. Experimental results show that SqueezeNet with Group Convolution achieved up to 90% accuracy, outperforming the original model. Additionally, the model exhibited significantly improved sensitivity to the minority class, indicating better performance under imbalanced conditions. These findings suggest that Group Convolution enhances not only computational efficiency but also classification capability. Therefore, this technique is applicable in developing automated diagnostic systems. Future research is encouraged to combine Group Convolution with methods such as attention mechanisms to achieve more optimal and reliable classification results. |
Description: | 12 Halaman |
URI: | https://repositori.uma.ac.id/handle/123456789/27662 |
Appears in Collections: | SP - Informatic Engineering |
Files in This Item:
File | Description | Size | Format | |
---|---|---|---|---|
208160003 - William Gultom - Fulltext.pdf | Fulltext | 788.75 kB | Adobe PDF | View/Open |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.