ESTRO 2024 - Abstract Book

S5046

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

Purpose/Objective:

To verify potential applicability for prostate cancer self-adaptive radiotherapy on 1.5T MR linear accelerator (MR LINAC) by exploring radiomics based online risk stratification.

Material/Methods:

This retrospective study was approved by the Hospital research ethics committee, with waiver of informed consent. One hundred and seventy-six consecutive PC patients who underwent MRgRT treatment on a 1.5T MR LINAC (Unity, Elekta) were retrospectively enrolled. Patients with MRI contraindications, prior prostate surgery/irradiation, history of other cancers, incomplete MRI scans, or lack of risk stratification were excluded. Each patient underwent planning CT followed by MRI on the 1.5T MRI-LINAC before any radiation, using a standardized 3D T2W-TSE sequence under consistent patient positioning and bladder protocols. An experienced radiation oncologist segmented the clinical target volume (CTV) on the planning CT with reference to the rigidly co-registered MRI, focusing anatomical matching of the prostate and seminal vesicles. Radiomics features were extracted from both the planning CT and MRI. A total of 1037 radiomic features from the PTV in both MR and CT images were extracted by using PyRadiomics v.2.2.0, including 93 original features, compliant with Imaging Biomarkers Standardization Initiative (IBSI), and 944 transformed features in wavelet and Laplacian-of-Gaussian domains. Highly correlated features were reduced using Spearman correlation (threshold 0.75). The endpoint was distinguishing D'Amico high-risk from non high-risk levels. Four classification models were applied to avoid data bias: support vector machine (SVM), multilayer perceptron (MLP), random forest (RF), and least absolute shrinkage and selection operator (LASSO). Features were balanced for high-risk and non-high-risk levels. Five-fold cross-validation (randomly 200 times per model) on the whole data set were introduced for the model training and evaluation. In the end, specificity, sensitivity, positive predictive value (PPV) and negative predictive value (NPV) were utilized for comprehensive evaluation of each multimodality model.

Results:

Fifty out of 176 consecutive male patients fulfilled all inclusion and exclusion criteria, and were included for analysis. In general, SVM and MLP performed well in terms of prediction accuracy (88.53% for SVM and 92.42% for MLP), sensitivity (90.56% for SVM and 96.75% for MLP) and specificity (86.66% for SVM and 87.78% for MLP) in the 5-fold cross validation. In contrast, RF exhibited the least favorable performance among the four model, with all evaluation metrics falling below 81%. Notably, MLP excelled in many aspects, achieving highest specificity, sensitivity, PPV and accuracy (87.78%, 96.75%, 96.72%, 92.42%).

Table 1. Evaluation matrix for multimodality models in 5-fold cross-validation

Models

Specificity

Sensitivity

PPV

NPV

Accuracy

SVM

86.66%

90.56%

79.35%

81.77%

88.53%

RF

70.87%

80.75%

79.18%

76.81%

75.84%

MLP

87.78%

96.75%

96.72%

90.73%

92.42%

LASSO

78.34%

91.89%

81.65%

92.40%

84.84%