0,94).Kedua, untuk mengatasi masalah transparansi, penelitian ini secara sistematis mengevaluasi empat teknik visual eXplainable AI (XAI). Metode Grad-CAM terbukti paling unggul dalam menjelaskan keputusan model, dengan mencapai nilai insertion area under curve (IAUC) tertinggi 0,9743 ± 0,0250 dan deletion area under curve (DAUC) terendah 0,0012 ± 0,0009. Keunggulan ini divalidasi oleh ahli radiologi dengan penilaian 4,8/5, dan analisis korelasi menunjukkan hubungan positif yang kuat (koefisien Spearman ? ? 0, 71) antara penilaian ahli dan metrik IAUC, yang memvalidasi relevansi klinis dari metrik yang digunakan.Ketiga, membangun sebuah kerangka kerja evaluasi yang baru dan komprehensif untuk mengukur sistem AI medis. Kerangka ini mengatasi keterbatasan pada penelitian sebelumnya dengan mengintegrasikan metrik kinerja teknis, interpretabilitas kuantitatif, dan validasi ahli klinis dalam satu penilaian terpadu. Kerangka evaluasi ini berhasil mengidentifikasi tiga konfigurasi sistem optimal untuk berbagai konteks implementasi.Secara keseluruhan, penelitian ini berhasil mengembangkan sistem segmentasi tumor otak yang memenuhi kebutuhan implementasi klinis dari segi efisiensi komputasi, akurasi diagnostik, dan transparansi pengambilan keputusan, memberikan landasan kuat untuk pengembangan sistem AI medis yang lebih dapat dipercaya dan praktis. Brain tumors represent a global health challenge that demands rapid and accurate diagnosis. However, the implementation of deep learning models in clinical practice faces several obstacles. Previous studies on brain tumor segmentation have often sacrificed image resolution to achieve computational efficiency when processing complex 3D MRI data, thereby risking the loss of critical details. In addition, the “black box” nature of these models hinders user trust, while existing evaluation frameworks frequently lack standardized interpretability metrics and integrated clinical validation.To address these challenges, this study makes three key contributions. First, it proposes a lightweight 3D ResU-Net architecture designed to efficiently process full-resolution MRI data. Through an optimized implementation, the proposed ResMobileU-Net model reduces the number of parameters by 50.5% (to 2.46 million) and computational complexity (FLOPs) by 45.3% (to 631.8 GFLOPs). Meanwhile, 3D ResGhostU-Net successfully reduces parameters by up to 58.6% compared to the baseline 3D ResU-Net architecture, while maintaining high segmentation accuracy (Dice coefficient > 0.94).Second, to enhance model transparency, this study systematically evaluates four visual Explainable AI (XAI) techniques. Grad-CAM is shown to be the most effective in explaining model decisions, achieving the highest insertion area under the curve (IAUC) of 0.9743 ± 0.0250 and the lowest deletion area under the curve (DAUC) of 0.0012 ± 0.0009. Its superiority is validated by a radiology expert who rated it 4.8 out of 5, and correlation analysis revealed a strong positive relationship (Spearman's ? ? 0, 71) between expert ratings and the IAUC metric, confirming the clinical relevance of the evaluation approach.Third, this work introduces a novel and comprehensive evaluation framework for assessing medical AI systems. The framework addresses limitations of prior studies by integrating technical performance metrics, quantitative interpretability scores, and expert clinical validation into a unified assessment. Using this framework, three optimal system configurations were identified for different implementation contexts.Overall, this study presents a brain tumor segmentation system that aligns with clinical implementation needs in terms of computational efficiency, diagnostic accuracy, and decision-making transparency, establishing a strong foundation for the development of more trustworthy and practical medical AI systems."> 0,94).Kedua, untuk mengatasi masalah transparansi, penelitian ini secara sistematis mengevaluasi empat teknik visual eXplainable AI (XAI). Metode Grad-CAM terbukti paling unggul dalam menjelaskan keputusan model, dengan mencapai nilai insertion area under curve (IAUC) tertinggi 0,9743 ± 0,0250 dan deletion area under curve (DAUC) terendah 0,0012 ± 0,0009. Keunggulan ini divalidasi oleh ahli radiologi dengan penilaian 4,8/5, dan analisis korelasi menunjukkan hubungan positif yang kuat (koefisien Spearman ? ? 0, 71) antara penilaian ahli dan metrik IAUC, yang memvalidasi relevansi klinis dari metrik yang digunakan.Ketiga, membangun sebuah kerangka kerja evaluasi yang baru dan komprehensif untuk mengukur sistem AI medis. Kerangka ini mengatasi keterbatasan pada penelitian sebelumnya dengan mengintegrasikan metrik kinerja teknis, interpretabilitas kuantitatif, dan validasi ahli klinis dalam satu penilaian terpadu. Kerangka evaluasi ini berhasil mengidentifikasi tiga konfigurasi sistem optimal untuk berbagai konteks implementasi.Secara keseluruhan, penelitian ini berhasil mengembangkan sistem segmentasi tumor otak yang memenuhi kebutuhan implementasi klinis dari segi efisiensi komputasi, akurasi diagnostik, dan transparansi pengambilan keputusan, memberikan landasan kuat untuk pengembangan sistem AI medis yang lebih dapat dipercaya dan praktis. Brain tumors represent a global health challenge that demands rapid and accurate diagnosis. However, the implementation of deep learning models in clinical practice faces several obstacles. Previous studies on brain tumor segmentation have often sacrificed image resolution to achieve computational efficiency when processing complex 3D MRI data, thereby risking the loss of critical details. In addition, the “black box” nature of these models hinders user trust, while existing evaluation frameworks frequently lack standardized interpretability metrics and integrated clinical validation.To address these challenges, this study makes three key contributions. First, it proposes a lightweight 3D ResU-Net architecture designed to efficiently process full-resolution MRI data. Through an optimized implementation, the proposed ResMobileU-Net model reduces the number of parameters by 50.5% (to 2.46 million) and computational complexity (FLOPs) by 45.3% (to 631.8 GFLOPs). Meanwhile, 3D ResGhostU-Net successfully reduces parameters by up to 58.6% compared to the baseline 3D ResU-Net architecture, while maintaining high segmentation accuracy (Dice coefficient > 0.94).Second, to enhance model transparency, this study systematically evaluates four visual Explainable AI (XAI) techniques. Grad-CAM is shown to be the most effective in explaining model decisions, achieving the highest insertion area under the curve (IAUC) of 0.9743 ± 0.0250 and the lowest deletion area under the curve (DAUC) of 0.0012 ± 0.0009. Its superiority is validated by a radiology expert who rated it 4.8 out of 5, and correlation analysis revealed a strong positive relationship (Spearman's ? ? 0, 71) between expert ratings and the IAUC metric, confirming the clinical relevance of the evaluation approach.Third, this work introduces a novel and comprehensive evaluation framework for assessing medical AI systems. The framework addresses limitations of prior studies by integrating technical performance metrics, quantitative interpretability scores, and expert clinical validation into a unified assessment. Using this framework, three optimal system configurations were identified for different implementation contexts.Overall, this study presents a brain tumor segmentation system that aligns with clinical implementation needs in terms of computational efficiency, diagnostic accuracy, and decision-making transparency, establishing a strong foundation for the development of more trustworthy and practical medical AI systems.">
Laporkan Masalah

Pengembangan Arsitektur Lightweight 3D ResU-Net untuk Segmentasi Tumor Otak Efisien dan Dapat Dijelaskan Melalui Validasi Multidimensi

DIAN NOVA KUSUMA HARDANI, Prof. Ir. Hanung Adi Nugroho, S.T., M.Eng., Ph.D., IPM., SMIEEE.; Dr.Eng. Ir. Igi Ardiyanto, S.T., M.Eng., IPM., SMIEEE.

2025 | Disertasi | S3 Teknik Elektro

Tumor otak merupakan tantangan kesehatan global yang memerlukan diagnosis cepat dan akurat. Namun, implementasi model deep learning dalam praktik klinis menghadapi beberapa tantangan. Penelitian sebelumnya dalam segmentasi tumor otak seringkali mengorbankan resolusi citra untuk mencapai efisiensi komputasi pada data MRI 3D yang kompleks, sehingga berisiko menghilangkan detail penting. Selain itu, sifat "kotak hitam" (black box) pada model menghambat kepercayaan pengguna, sementara kerangka evaluasi yang ada seringkali tidak memiliki metrik interpretabilitas yang terstandarisasi dan validasi klinis yang terintegrasi.

Menjawab tantangan tersebut, penelitian ini memberikan tiga kontribusi utama. Pertama, mengembangkan arsitektur lightweight 3D ResU-Net yang efisien untuk memproses data MRI resolusi penuh. Implementasi optimal melalui model ResMobileU-Net berhasil mengurangi jumlah parameter sebesar 50,5% (menjadi 2,46 juta) dan kompleksitas komputasi (FLOPs) sebesar 45,3% (menjadi 631,8 GFLOPs). Sementara itu, 3D ResGhostU-Net berhasil menekan parameter hingga 58,6% dibandingkan arsitektur 3D ResU-Net baseline, dengan tetap mempertahankan akurasi segmentasi yang tinggi (koefisien Dice > 0,94).

Kedua, untuk mengatasi masalah transparansi, penelitian ini secara sistematis mengevaluasi empat teknik visual eXplainable AI (XAI). Metode Grad-CAM terbukti paling unggul dalam menjelaskan keputusan model, dengan mencapai nilai insertion area under curve (IAUC) tertinggi 0,9743 ± 0,0250 dan deletion area under curve (DAUC) terendah 0,0012 ± 0,0009. Keunggulan ini divalidasi oleh ahli radiologi dengan penilaian 4,8/5, dan analisis korelasi menunjukkan hubungan positif yang kuat (koefisien Spearman ? ? 0, 71) antara penilaian ahli dan metrik IAUC, yang memvalidasi relevansi klinis dari metrik yang digunakan.

Ketiga, membangun sebuah kerangka kerja evaluasi yang baru dan komprehensif untuk mengukur sistem AI medis. Kerangka ini mengatasi keterbatasan pada penelitian sebelumnya dengan mengintegrasikan metrik kinerja teknis, interpretabilitas kuantitatif, dan validasi ahli klinis dalam satu penilaian terpadu. Kerangka evaluasi ini berhasil mengidentifikasi tiga konfigurasi sistem optimal untuk berbagai konteks implementasi.

Secara keseluruhan, penelitian ini berhasil mengembangkan sistem segmentasi tumor otak yang memenuhi kebutuhan implementasi klinis dari segi efisiensi komputasi, akurasi diagnostik, dan transparansi pengambilan keputusan, memberikan landasan kuat untuk pengembangan sistem AI medis yang lebih dapat dipercaya dan praktis.

Brain tumors represent a global health challenge that demands rapid and accurate diagnosis. However, the implementation of deep learning models in clinical practice faces several obstacles. Previous studies on brain tumor segmentation have often sacrificed image resolution to achieve computational efficiency when processing complex 3D MRI data, thereby risking the loss of critical details. In addition, the “black box” nature of these models hinders user trust, while existing evaluation frameworks frequently lack standardized interpretability metrics and integrated clinical validation.

To address these challenges, this study makes three key contributions. First, it proposes a lightweight 3D ResU-Net architecture designed to efficiently process full-resolution MRI data. Through an optimized implementation, the proposed ResMobileU-Net model reduces the number of parameters by 50.5% (to 2.46 million) and computational complexity (FLOPs) by 45.3% (to 631.8 GFLOPs). Meanwhile, 3D ResGhostU-Net successfully reduces parameters by up to 58.6% compared to the baseline 3D ResU-Net architecture, while maintaining high segmentation accuracy (Dice coefficient > 0.94).

Second, to enhance model transparency, this study systematically evaluates four visual Explainable AI (XAI) techniques. Grad-CAM is shown to be the most effective in explaining model decisions, achieving the highest insertion area under the curve (IAUC) of 0.9743 ± 0.0250 and the lowest deletion area under the curve (DAUC) of 0.0012 ± 0.0009. Its superiority is validated by a radiology expert who rated it 4.8 out of 5, and correlation analysis revealed a strong positive relationship (Spearman's ? ? 0, 71) between expert ratings and the IAUC metric, confirming the clinical relevance of the evaluation approach.

Third, this work introduces a novel and comprehensive evaluation framework for assessing medical AI systems. The framework addresses limitations of prior studies by integrating technical performance metrics, quantitative interpretability scores, and expert clinical validation into a unified assessment. Using this framework, three optimal system configurations were identified for different implementation contexts.

Overall, this study presents a brain tumor segmentation system that aligns with clinical implementation needs in terms of computational efficiency, diagnostic accuracy, and decision-making transparency, establishing a strong foundation for the development of more trustworthy and practical medical AI systems.

Kata Kunci : Segmentasi tumor otak, deep learning, explainable AI, MRI 3D, lightweight 3D ResU-Net, interpretabilitas visual, validasi klinis.

  1. S3-2025-468290-abstract.pdf  
  2. S3-2025-468290-bibliography.pdf  
  3. S3-2025-468290-tableofcontent.pdf  
  4. S3-2025-468290-title.pdf