Radiation medicine and protection · 2026
A comparison of dimensionality reduction approaches for pre-treatment PSMA-PET/CT radiomics in prostate adenocarcinoma outcome prediction.
СтатьяНазвание и аннотация: Перевод пока не подготовлен
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Singh Apurva, Mendes William Silva, Oh Sang-Bo, Guler Ozan Cem, Elmali Aysenur, Demirhan Birhan, Sawant Amit, Tran Phuoc, Onal Cem, Ren Lei
Аннотация
OBJECTIVE: To compare dimensionality-reduction methods for building prognostic models predicting metastasis-free survival (MFS) in localized prostate adenocarcinoma (PCa) patients treated with androgen-deprivation therapy and external radiotherapy using clinical factors and prostate-specific membrane antigen (PSMA)-PET/CT radiomics from primary tumor and nodal volumes.
METHODS: A total of 134 localized PCa patients (28 with nodal involvement) were analyzed. Gross tumor volumes for primary tumors (GTVp) and nodes (GTVn) were segmented on CT and PET scans; a 5-mm peritumoral ring was defined. Radiomic features were normalized and reduced using three techniques: principal component analysis (PCA), supervised, and unsupervised feature selection. Model 1 combined tumor and nodal radiomics via volume-weighted averaging and consisted of 12 predictors including clinical variables (age, Gleason score, initial PSA, PSA relapse) and radiomics from primary, nodal, and ring regions. Data imbalance (24 metastasis, 110 no metastasis) was addressed using a 70:30 train-test split with imbalance correction applied to train set. Univariate Cox regression (P < 0.05) identified top predictors from train set; multivariate Cox regression was performed on corrected training data and applied to test data. Model 2 used clinical variables and radiomics from GTVp + ring; Model 3 used clinical data alone. Binary classification for five-year MFS was also evaluated.
RESULTS: Supervised feature selection achieved highest performance. Model 1 test had c-score 0.71 (0.65-0.72). The five-year MFS test classification was sensitivity 80.1 %, specificity 85.4 %, and AUC 0.84. Unsupervised and PCA methods showed slightly lower results (test c-scores: 0.70 and 0.69, respectively). Model 1 consistently outperformed Model 2 with c-score 0.64 and AUC 0.79, as well as Model 3 with c-score 0.54 and AUC 0.68 across all dimensionality-reduction techniques.
CONCLUSION: Supervised feature selection yielded the highest c-scores and AUCs for the models. Integrating PSMA-PET/CT radiomics from primary, nodal, and peritumoral regions with clinical factors significantly improved MFS prediction, highlighting multi-regional radiomics as a promising biomarker for personalized therapy in prostate cancer.
METHODS: A total of 134 localized PCa patients (28 with nodal involvement) were analyzed. Gross tumor volumes for primary tumors (GTVp) and nodes (GTVn) were segmented on CT and PET scans; a 5-mm peritumoral ring was defined. Radiomic features were normalized and reduced using three techniques: principal component analysis (PCA), supervised, and unsupervised feature selection. Model 1 combined tumor and nodal radiomics via volume-weighted averaging and consisted of 12 predictors including clinical variables (age, Gleason score, initial PSA, PSA relapse) and radiomics from primary, nodal, and ring regions. Data imbalance (24 metastasis, 110 no metastasis) was addressed using a 70:30 train-test split with imbalance correction applied to train set. Univariate Cox regression (P < 0.05) identified top predictors from train set; multivariate Cox regression was performed on corrected training data and applied to test data. Model 2 used clinical variables and radiomics from GTVp + ring; Model 3 used clinical data alone. Binary classification for five-year MFS was also evaluated.
RESULTS: Supervised feature selection achieved highest performance. Model 1 test had c-score 0.71 (0.65-0.72). The five-year MFS test classification was sensitivity 80.1 %, specificity 85.4 %, and AUC 0.84. Unsupervised and PCA methods showed slightly lower results (test c-scores: 0.70 and 0.69, respectively). Model 1 consistently outperformed Model 2 with c-score 0.64 and AUC 0.79, as well as Model 3 with c-score 0.54 and AUC 0.68 across all dimensionality-reduction techniques.
CONCLUSION: Supervised feature selection yielded the highest c-scores and AUCs for the models. Integrating PSMA-PET/CT radiomics from primary, nodal, and peritumoral regions with clinical factors significantly improved MFS prediction, highlighting multi-regional radiomics as a promising biomarker for personalized therapy in prostate cancer.