Abstract
527
Objectives: Reliable classification of prostate cancer (PCa) lesions from 18F-DCFPyL prostate-specific membrane antigen (PSMA) PET images is an important clinical need for the diagnosis and prognosis of PCa (1). For this purpose, a PSMA-reporting and data systems (RADS) was developed to classify PSMA-targeted PET scans into categorizations that reflect the likelihood of PCa (1). Deep learning-based methods have recently shown promise in the classification tasks in medical images (2,3). In this work, we aimed to develop an automated deep learning-based method for lesion classification in 18F-DCFPyL PET images.
Methods: 18F-DCFPyL PSMA PET images of 267 patients with PCa were manually segmented by four nuclear medicine physicians. Each segmented region was assigned to one of nine possible PSMA-RADS categories (Table 1). The dataset contained 3,724 PCa lesions where each patient had approximately 14 lesions on average. A deep 3D convolutional neural network (3D-CNN) was developed to classify those lesions (Fig. 1a). The 3D PET images were cropped to yield cubic volumes-of-interest (VOIs) around the center of each lesion as the input to the network. VOI sizes varying from 83, 163, 323, 643, and 1283 cubic voxels were investigated. The VOI size that resulted in the best overall accuracy on the validation set was used in the final model. The VOI containing the lesion and the corresponding manual segmentation were given as inputs to the network. The output of the 3D-CNN was the predicted PSMA-RADS score. The 3,724 PCa lesions were randomly partitioned into training, validation and test datasets containing 2,607, 559 and 558 lesions, respectively, using a 70%/15%/15% split. The network was trained with a cross-entropy loss function and a first-order stochastic gradient-based optimization algorithm (4). Early stopping based on monitoring the error on the validation set was applied to prevent overfitting during training (5). The proposed method was then evaluated on the test set by assessing standard evaluation metrics, including overall accuracy, confusion matrix and area under receiver operating characteristic curve (AUROC). Overall accuracy was defined as the number of correctly classified observations divided by the total number of observations.
Results: Using an input size of 1283 cubic voxels, the proposed method yielded an overall accuracy of 67.3% (95% confidence interval (CI): 63.4%, 71.2%)) and AUROCs of 0.97, 0.94, 0.92, 0.90, 0.76, 0.96, 0.99, 0.89 and 0.89 for PSMA-RADS-1A, PSMA-RADS-1B, PSMA-RADS-2, PSMA-RADS-3A, PSMA-RADS-3B, PSMA-RADS-3C, PSMA-RADS-3D, PSMA-RADS-4 and PSMA-RADS-5, respectively, on the validation set (Fig. 1b). The input VOI size of 1283 cubic voxels significantly outperformed the smaller VOI sizes of 83, 163, 323 (paired sample t-test p-value<0.05) on the basis of overall accuracy (Fig. 1c). Using an input size of 1283 cubic voxels, the proposed method yielded an overall accuracy of 67.4% (95% CI: 63.5%, 71.3%) and AUROCs of 0.93, 0.95, 0.89, 0.93, 0.82, 0.98, 0.96, 0.88 and 0.90 for PSMA-RADS-1A, PSMA-RADS-1B, PSMA-RADS-2, PSMA-RADS-3A, PSMA-RADS-3B, PSMA-RADS-3C, PSMA-RADS-3D, PSMA-RADS-4 and PSMA-RADS-5, respectively, on the test set (Fig. 1d). The confusion matrix for the proposed method on the test set is shown in Fig. 1e.
Conclusions: A deep learning-based approach for lesion classification in PSMA PET images was developed and showed significant promise towards automated classification of PCa lesions.
