Evaluating the Effect of Multilabel and Single Label Models on Prostate Cancer Lesion Segmentation in Ga-68 PSMA-11 PET/CT

Introduction: The accurate segmentation of lesions on a precision radiomolecular imaging like the Ga-68 PSMA-11 PET/CT is essential for effective and personalized treatment planning in prostate cancer (PCa). Currently, lesion delineation is typically performed manually, but this method is time-consuming, error-prone, and can suffer from intra- and inter-reader variability. Automated quantitation using deep learning algorithms for segmentation has the potential to improve feature extraction, tumor staging, radiotherapy planning, and treatment response assessment. Challenges in developing automatic algorithms for tumor segmentation include the poor resolution and high statistical noise in PET images, uptake of tracers in both normal and tumor tissue, time-dependent blood pool signal, and variability in the inter-subject uptake and data acquisition. Most current deep learning methods only use models for tumor lesion segmentation. We propose that segmenting normal tissues in addition will enhance segmentation performance. Therefore, in this study, we compare the performance of multilabel and single label models for lesion segmentation and detection in Ga-68 PSMA-11 PET/CT scans.

Methods: For this study, Ga-68 PSMA-11 PET/CT images of 50 patients with PCa were analyzed. Trained annotators manually segmented the images, which were then validated by expert physicians. The annotators were instructed to segment multiple regions of interest (ROIs), including the liver, kidney, spleen, bladder, salivary glands, lacrimal glands, nasopharyngeal mucosa, tumor lesions, and injection site. The dataset contains 2233 lesions in total, of which 304 are in the held-out dataset. A multi-label and single label deep convolutional neural network (CNN) were implemented to identify PCa lesions as well as the multiple ROIs. The 50 patient images were randomly divided into a training set with 40 images and a held-out test set with 10 images. For each model, the hyperparameters of the network were optimized using a 4-fold cross-validation on the training set. The network was trained with a cross-entropy and dice loss function, and the ADAM optimization algorithm. The proposed methods were then evaluated on the test set for each fold. The segmentation accuracy was assessed using standard evaluation metrics, including the Dice Similarity Coefficient (DSC) and Positive Predictive Value (PPV). The DSC is a measure of overlap, with higher values indicating more accurate segmentation. PPV is the probability that a positive result is actually correct.

Results: The proposed fully automated deep-learning method for multi-label PCa lesion detection yielded a DSC, PPV, of 0.7235 (95% confidence interval (CI): 0.6831, 0.7689), and 0.8059 (95% CI: 0.7275, 0.8844), respectively, on the held-out test set. For comparison we also tested single label model on the same held out test dataset which yielded a DSC, PPV of 0.3516 (95% CI: 0.1551, 0.5481), and 0.7489 (95% CI: 0.7051, 0.7926), respectively. Multilabel model yielded 0.9471 (95% CI: 0.9426, 0.9516), 0.9342 (95% CI: 0.9251, 0.9428), 0.9206 (95% CI: 0.9122, 0.9290), 0.9126 (95% CI: 0.8990, 0.9261), 0.8846 (95% CI: 0.8280, 0.9411), 0.8515 (95% CI: 0.7874, 0.9157), 0.8080 (95% CI: 0.7595, 0.8565), and 0.8460 (95% CI: 0.7064, 0.9856) for liver, kidney, bladder, spleen, salivary glands, lacrimal glands, nasopharyngeal mucosa and injection site, respectively.

Conclusions: Our study demonstrated the effectiveness of using a fully automated deep-learning method with multiple labels for detecting prostate cancer lesions on Ga-68 PSMA-11 PET/CT. We found that this approach outperformed a single label model in terms of both DSC and PPV. This suggests that by identifying normal uptake regions in addition to tumor lesions, the multilabel model can improve its detection and segmentation performance. Overall, our results highlight the potential of deep learning approaches for accurately identifying prostate cancer lesions.