Tesis

Penerapan Deep Learning untuk Auto-Segmentasi Radioterapi Kanker Serviks : Training, Validasi, dan Implementasi Klinis = Application of Deep Learning for Cervical Cancer Radiotherapy Auto-Segmentation: Training, Validation, and Clinical Implementation.

Pendahuluan : Proses delineasi clinical target volume (CTV) dan organ at risk (OAR) pada radioterapi kanker serviks merupakan tahapan krusial namun memerlukan waktu dan rentan variasi antar-operator. Penelitian ini mengevaluasi performa autokontur berbasis AI (PVMed) untuk struktur CTV Pelvis, kandung kemih, dan rektum pada kanker serviks stadium lokal lanjut. Tujuan : Menilai validasi kuantitatif dan kualitatif performa autokontur AI PVMed dibandingkan kontur manual pada kasus kanker serviks stadium IIB–IIIC1 (FIGO 2018). Metode : Desain kohort retrospektif menggunakan 100 dataset struktur RTDICOM kanker serviks di IPTOR RSCM. Sebanyak 70 dataset untuk training dan 30 dataset untuk testing, dengan tambahan 20 dataset (10 training + 10 testing) untuk validasi. Evaluasi kuantitatif dilakukan pada dataset testing dengan metrik Dice Similarity Coefficient (DSC) dan Hausdorff Distance (HD). Evaluasi kualitatif dilakukan oleh 2 reviewer konsultan abdominopelvis pada 60 dataset (30 autokontur AI dan 30 kontur manual peneliti) menggunakan daftar error mayor/minor, serta penilaian penerimaan klinis berbasis protokol GHG yang dimodifikasi. Hasil : Kinerja geometrik pada testing stabil pada tiga skenario jumlah dataset training. Median HD berturut-turut pada model training 30, 50, dan 70 dataset adalah 0,87; 0,88; dan 0,89. Sedangkan nilai DSC dengan nilai 3,49; 2,87; dan 2,74. Secara klinis, tingkat penerimaan (skor tertinggi = “3”) untuk autokontur kandung kemih mencapai 73,3% pada kedua reviewer; rektum 50,0% (reviewer 1) dan 33,3% (reviewer 2); sedangkan CTV Pelvis rendah (20,0% dan 23,3%), dengan mayoritas dinilai tidak dapat diterima (skor “1”). Kesimpulan : Autokontur AI PVMed berpotensi digunakan pada delineasi kanker serviks stadium lokal lanjut, dengan performa terbaik pada kandung kemih (umumnya tanpa revisi) dan rektum (memerlukan revisi minor). Namun, CTV Pelvis masih memerlukan revisi bermakna, terutama pada batas atas delineasi.
Kata kunci : Kanker serviks, Radioterapi, Autokontur, PVMed, Deep learning, DSC, HD


Introduction : Delineation of the clinical target volume (CTV) and organs at risk (OAR) in cervical cancer radiotherapy is a critical step; however, it is timeconsuming and prone to inter-operator variability. This study evaluated the performance of an AI-based autocontouring system (PVMed) for pelvic CTV, bladder, and rectum structures in locally advanced cervical cancer. Objective: To quantitatively and qualitatively validate the performance of PVMed AI autocontouring compared with manual contouring in cervical cancer cases staged IIB–IIIC1 (FIGO 2018). Methods: A retrospective cohort design was conducted using 100 RT-DICOM datasets of cervical cancer from IPTOR RSCM.As many as 70 datasets were used for training and 30 for testing, with an additional 20 datasets (10 training + 10 testing) used for validation. Quantitative evaluation was performed on the testing dataset using the Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD). Qualitative evaluation was conducted by two abdominopelvic radiation oncology consultants on 60 datasets (30 AI autocontours and 30 manual contours created by the researcher) using a checklist of major/minor errors, as well as a clinical acceptability assessment based on a modified GHG protocol. Results: Geometric performance on the testing set remained stable across three training dataset size scenarios. The median HD values for models trained on 30, 50, and 70 datasets were 0.87; 0.88; and 0,89. The value of DSC were 3,49; 2,87; dan 2,74. Clinically, the acceptability rate (highest score = “3”) for bladder autocontours reached 73.3% for both reviewers; rectum autocontours achieved 50.0% (reviewer 1) and 33.3% (reviewer 2); whereas pelvic CTV acceptability was low (20.0% and 23.3%), with the majority rated as not acceptable (score “1”). Conclusion: PVMed AI autocontouring shows potential for use in locally advanced cervical cancer radiotherapy delineation, with the best performance observed for the bladder (generally requiring no revision) and rectum (requiring minor revisions). However, pelvic CTV still requires substantial revisions, particularly at the superior boundary of delineation.
Keywords: Cervical cancer, radiotherapy, autocontouring, PVMed, deep learning, DSC, HD.