Prediksi Dosis Radioterapi Kanker Nasofaring Berbasis Transfer Learning HD U-Net MONAI: Adaptasi Dari Dataset OpenKBP IMRT Ke Dataset VMAT Busur Ganda

Sururi, Isthar Bismuth (2026) Prediksi Dosis Radioterapi Kanker Nasofaring Berbasis Transfer Learning HD U-Net MONAI: Adaptasi Dari Dataset OpenKBP IMRT Ke Dataset VMAT Busur Ganda. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5001221059-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Radioterapi merupakan modalitas utama pengobatan kanker nasofaring (KNF), dengan Volumetric-Modulated Arc Therapy (VMAT) sebagai standar klinis modern menggantikan Intensity-Modulated Radiotherapy (IMRT). Model deep learning berbasis Hierarchically Densely Connected U-Net (HD U-Net) untuk prediksi dosis umumnya dilatih pada dataset publik OpenKBP (berbasis IMRT), namun seringkali mengalami penurunan performa signifikan saat diterapkan pada VMAT busur ganda akibat perbedaan mekanika pengiriman dosis. Penelitian ini menganalisis efektivitas cross-technique transfer learning dari domain OpenKBP (240 pasien IMRT) ke domain klinis VMAT busur ganda (51 pasien KNF di RSUD dr. Mohamad Soewandhie) menggunakan HD U-Net berbasis PyTorch dan MONAI. Data RS dikonversi ke format matriks sparse OpenKBP melalui resampling isotropik, normalisasi HU, dan mapping nama ROI. Model pre-trained (30 epoch OpenKBP) di-fine-tune selama 150 epoch pada 40 pasien RSUD Soewandhie dengan strategi differential learning rate, dan dievaluasi menggunakan Average Percent Prediction Error (APPE) terhadap dosis 70 Gy pada 5 pasien uji. Hasil menunjukkan model pre-trained tanpa adaptasi menghasilkan APPE D_mean PTV 36,07 ± 1,95% dan APPE D_max OAR 29,68 ± 11,00%, membuktikan domain gap substansial antara IMRT dan VMAT busur ganda. Setelah fine-tuning, APPE D_mean PTV turun menjadi 0,23 ± 0,06% dan APPE D_max OAR menjadi 0,41 ± 0,27% , melampaui hasil literatur acuan. Analisis DVH dan dose color-wash mengonfirmasi model fine-tuned mereproduksi pola dosis konformal VMAT busur ganda secara akurat, membuktikan transfer learning lintas teknik IMRT ke VMAT efektif dan layak dikembangkan untuk penelitian selanjutnya.
=================================================================================================================================
Radiotherapy is the primary treatment modality for nasopharyngeal carcinoma (NPC), with volumetric-modulated arc therapy (VMAT) serving as the modern clinical standard, replacing intensity-modulated radiotherapy (IMRT). Deep learning models based on the hierarchically densely connected U-Net (HD U-Net) for dose prediction are commonly trained on the public OpenKBP dataset (IMRT-based); however, they often experience significant performance degradation when applied to VMAT double arcs owing to fundamental differences in dose delivery mechanics. This study analyzed the effectiveness of cross-technique transfer learning from the OpenKBP domain (240 IMRT patients) to a local clinical VMAT domain (51 NPC patients at RSUD dr. Mohamad Soewandhie) using a PyTorch- and Medical Open Network for AI-based HD U-Net. Local data were converted into the OpenKBP sparse matrix format through isotropic resampling, HU normalization, and ROI name mapping, validated by DVH agreement with the TPS ground truth. The pre-trained model (30 epochs on OpenKBP) was fine-tuned for 150 epochs on 40 local patients using a differential learning rate strategy and evaluated using the relative APPE to a 70 Gy prescription dose on five test patients. The results showed that the pre-trained model without adaptation produced a PTV D_mean APPE of 36.07±1.95% and OAR D_max APPE of 29.68 ± 11.00%, demonstrating a substantial domain gap between IMRT and dual-arc VMAT. After fine-tuning, PTV D_mean APPE decreased to 0.23 ± 0.06% and OAR D_max APPE to 0.41 ± 0.27%, surpassing the reference literature. DVH and dose color-wash analyses confirmed that the fine-tuned model accurately reproduces the conformal dose distribution pattern of dual-arc VMAT, demonstrating that cross-technique transfer learning from IMRT to VMAT is effective and warrants further development in future research.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	IMRT, Kanker Nasofaring, Prediksi Dosis, Transfer Learning, VMAT Busur Ganda, Dose Prediction, IMRT, Nasopharyngeal Carcinoma, Transfer Learning, VMAT Double Arc
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QC Physics > QC795.5 Radioactivity and radioactive Instruments and apparatus (General)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Physics > 45201-(S1) Undergraduate Thesis
Depositing User:	Isthar Bismuth Sururi
Date Deposited:	27 Jul 2026 07:52
Last Modified:	27 Jul 2026 07:52
URI:	http://repository.its.ac.id/id/eprint/137986

Actions (login required)

View Item