Abstract
2975
Introduction: To examine the usefulness of machine learning (ML) approaches employing pretreatment clinical and 2-deoxy-2-[18F]fluoro-D-glucose (18F-FDG) PET-based radiomic features for predicting disease recurrence in patients with breast cancers who received radical surgery.
Methods: This retrospective study included 121 patients with breast cancers (UICC stage 0 to III) who underwent 18F-FDG-PET/CT before radical surgery. Thirteen clinical features [T stage, N stage, M stage, UICC stage, histology, 3 hormone status (estrogen receptor, progesterone receptor and HER-2), Ki-67, molecular subtype and 3 tumor markers (CEA, CA153 and BCA225)] and 40 18F-FDG PET-based radiomic features [9 first order features, 6 features of gray-level co-occurrence matrix, 3 features of neighborhood gray-level different matrix, 11 features of gray-level run length matrix and 11 features of gray-level zone length matrix] were considered to predict disease recurrence using ML approaches. Five ML algorithms were used for binary classification (disease progression/non-progression): random forest (RF), neural network, k-nearest neighbors, logistic regression, and a support vector machine. Receiver operating characteristic curve analysis was performed to obtain the area under the curve (AUC). AUC, accuracy, F1 score, precision (positive predictive value), and recall (sensitivity) were the performance measures computed. Given that the small number of patients with disease recurrence in the training cohort might lead to a potential over-fitting, the training dataset was augmented by synthetic minority over-sampling technique (SMOTE). To minimize the negative influence of overfitting, 10-fold cross-validation was used. The data were stratified by event and randomly assigned into training (80%) and validation (20%) cohorts.
Results: Of 97 patients in the training cohort, 10 patients were classified as disease progression, and 87 as non-progression. The 24 patients in the validation cohort consisted of 3 disease progression and 21 non-progression cases.
In the training cohort, the augmentation by SMOTE resulted in 87 data for the patients with disease recurrence. Among 5 ML models, the RF model achieved the best performance classifier (RF: AUC = 0.992, accuracy = 0.977, F1 score = 0.977, precision = 0.978, and recall = 0.977), followed by the neural network model (AUC = 0.978, accuracy = 0.966, F1 score = 0.965, precision =0.968, and recall = 0.966). Other 3 ML models also showed good performance (AUC ranging from 0.951 to 0.973) in the training dataset.
In the validation cohort, the RF model showed the best-performing classifier with high diagnostic accuracy (AUC = 0.773, accuracy = 0.833, F1 score = 0.833, precision = 0.833, and recall = 0.833), followed by the logistic regression model (AUC = 0.723, accuracy = 0.875, F1 score = 0.866, precision = 0.900, and recall = 0.875), although the AUC of the two ML models decreased compared with those of the training sets. The other 3 ML models showed poor performances in the validation tests (AUC ranging from 0.364 to 0.614).
Conclusions: ML approaches employing clinical and pretreatment 18F-FDG PET-based radiomic features may be useful for predicting disease recurrence in breast cancer patients who recieved radical surgery.