Klasifikasi Tingkat Kecurigaan Kanker Prostat Berbasis PI-RADS menggunakan Deep Learning dengan Visualisasi Lesi pada Citra MRI Multiparametrik

Tarigan, Jefta Kenna Shalom (2026) Klasifikasi Tingkat Kecurigaan Kanker Prostat Berbasis PI-RADS menggunakan Deep Learning dengan Visualisasi Lesi pada Citra MRI Multiparametrik. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5049221018-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (7MB) | Request a copy

Abstract

Kanker prostat merupakan keganasan kedua tersering pada pria, dengan 13.130 kasus baru di Indonesia pada tahun 2022, sehingga diperlukan sistem diagnosis yang akurat sekaligus efisien. Salah satu modalitas utama diagnosis adalah multiparametric MRI (mpMRI), yang dinilai menggunakan sistem skoring standar Prostate Imaging-Reporting and Data System (PI-RADS). Namun, penilaian PI-RADS masih bergantung pada interpretasi subjektif radiolog sehingga menimbulkan variabilitas antar-pembaca yang signifikan. Penelitian ini bertujuan mengembangkan sistem Computer-Aided Diagnosis (CAD) berbasis deep learning dan radiomik untuk klasifikasi otomatis tingkat kecurigaan kanker prostat sesuai standar PI-RADS, dilengkapi visualisasi lesi. Penelitian retrospektif ini menggunakan sekuens mpMRI (T2-weighted, DWI, dan ADC) dari dua dataset publik multi-pusat, yaitu PI-CAI (n = 1.475) dan Prostate158 (n = 139), dengan anotasi ahli sebagai ground truth. Sistem dibangun melalui empat tahap: (1) segmentasi kelenjar prostat dan lesi intraprostatik menggunakan pendekatan dual nnU-Net; (2) ekstraksi fitur radiomik dengan pendekatan Tensor Radiomics (PyRadiomics); (3) seleksi fitur berbasis L1-regularized Logistic Regression; dan (4) klasifikasi bertingkat (cascade) tiga tahapmenggunakan ensemble tiga algoritma (XGBoost, LightGBM, dan Random Forest) dengan perutean keras dan ambang berorientasi sensitivitas. Interpretabilitas disediakan melalui dua lapis, yaitu tumor probability map untuk lokalisasi lesi dan SHapley Additive exPlanations (SHAP) untuk atribusi fitur. Pada tahap segmentasi, model mencapai Dice Similarity Coefficient (DSC) sebesar 0,908 untuk kelenjar prostat pada uji eksternal Prostate158, sedangkan performa segmentasi lesi bersifat bergantung pada tingkat kecurigaan (suspicion-dependent). Model klasifikasi cascade final mencapai macro-averaged AUROC sebesar 0,888 (95% CI: 0,797–0,959), dengan AUC per kelas PI-RADS 3, 4, dan 5 masing-masing 0,966, 0,841, dan 0,857, serta akurasi keseluruhan 0,75. Model radiomik bertingkat yang dikembangkan menunjukkan performa diskriminasi yang baik untuk grading tingkat kecurigaan PI-RADS, termasuk pada kelas PI-RADS 3 yang ekuivokal. Dipadukan dengan visualisasi lesi yang interpretatif, sistem CAD ini berpotensi menjadi alat bantu keputusan yang meningkatkan objektivitas diagnosis dan mengurangi variabilitas antar-radiolog dalam praktik klinis.
======================================================================================================================================
Prostate cancer is the second most common malignancy in men, with 13,130 new cases in Indonesia in 2022, underscoring the need for a diagnostic system that is accurate yet efficient. One of the principal diagnostic modalities is multiparametric MRI (mpMRI), which is assessed using the standardized Prostate Imaging-Reporting and Data System (PI-RADS). However, PI-RADS assessment still relies on the radiologist's subjective interpretation, resulting in substantial inter-reader variability. This study aims to develop a deep learning– and radiomics–based Computer-Aided Diagnosis (CAD) system for the automated classification of prostate cancer suspicion levels according to the PI-RADS standard, complemented by lesion visualization. This retrospective study used mpMRI sequences (T2-weighted, DWI, and ADC) from two multi-center public datasets, namely PI-CAI (n = 1,475) and Prostate158 (n = 139), with expert annotations as ground truth. The system was developed in four stages: (1) segmentation of the prostate gland and intraprostatic lesions using a dual nnU-Net approach; (2) radiomic feature extraction using a Tensor Radiomics approach (PyRadiomics); (3) feature selection based on L1-regularized Logistic Regression; and (4) a three-stage cascade classificationusing an ensemble of three algorithms (XGBoost, LightGBM, and Random Forest) with hard routing and sensitivity-oriented thresholds. Interpretability was provided through two complementary layers: a tumor probability map for lesion localization and SHapley Additive exPlanations (SHAP) for feature attribution. For segmentation, the model achieved a Dice Similarity Coefficient (DSC) of 0.908 for the prostate gland on external Prostate158 testing, whereas lesion segmentation performance was suspicion-dependent. The final cascade classification model achieved a macro-averaged AUROC of 0.888 (95% CI: 0.797–0.959), with per-class AUCs for PI-RADS 3, 4, and 5 of 0.966, 0.841, and 0.857, respectively, and an overall accuracy of 0.75. The proposed multi-stage radiomics model demonstrated strong discriminative performance for PI-RADS suspicion grading, including for the equivocal PI-RADS 3 class. Combined with interpretable lesion visualization, this CAD system is a promising decision-support tool that may improve diagnostic objectivity and reduce inter-radiologist variability in clinical practice., and serve as a clinical decision support tool that can enhance the quality of patient care

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Kanker Prostat, PI-RADS, nnU-Net, MRI multiparametrik, Radiomik, Prostate Cancer, Multiparametric MRI, Radiomics
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of medicine and health (MEDICS) > Medical Technology
Depositing User:	Jefta Kenna Shalom Tarigan
Date Deposited:	04 Aug 2026 04:13
Last Modified:	04 Aug 2026 04:13
URI:	http://repository.its.ac.id/id/eprint/139908

Actions (login required)

View Item