Multi-Task Deep Learning for the Detection of Lesions on 68Ga-PSMA PET/CT Imaging

Purpose: The emerging PSMA targeted radionuclide therapy provides an effective method for the treatment of advanced metastatic prostate cancer. To optimize diagostics, therapy monitoring and ultimately the theranostic benefit, it is urgently needed to characterize all the lesions to target before the treatment. However, this is extremely challenging considering the factor that dozens of lesions of heterogenous size and uptake may distribute in a variety of anatomical context with different background. Until now, there is no successful computer-aided lesion detection methods for PSMA imaging. Methods: A cohort of 71 patients with advanced metastatic prostate cancer were scanned with 68Ga-PSMA-11 PET/CT. For proof-of-concept, we focus on the detection and segmentation of bone & lymph node lesions in the pelvic area. To train the network, the bone and lymph node lesions were manually labelled by a nuclear medicine expert. A multi-task deep learning architecture (MulTi-Net) based on fully convolutional neural networks was developed to detect the lesions. It aimed to first extract salient features from PET and CT, the combined features would then be adopted to automatically detect all the lesions in a 3D manner. The framework contains five fully connected convolutional layers and two sigmoid classification layers. In contrast to conventional V-Net, it deleted pooling layer to avoid possible losing of fine texture information. A cross-hair filter was integrated to extensively reduce hyperparameters and speed up the training procedure. An additional regularization was added to deal with class imbalanced tumor-to-background ratio. For comparison, the detection accuracy of conventional W-Net (cascaded V-Nets) were calculated. Results: Compared with conventional W-Net, the multi-task deep learning has improved the detection precision from 72.8% to 90.2%, recall from 59.9% to 76.3% for bone lesion. For the lymph node lesion (n=63), it improved the detection precision from 57.8% to 81.4% ad recall from 44.1% to 62.6%. Conclusion: We proposed the first deep learning method for automatic detection of lesions on 68Ga-PSMA-11 PET/CT images. A multi-task deep learning method was developed to improve the detection accuracy compared with conventional W-Net. The preliminary test on pelvic area confirmed the potential of deep learning methods. At the moment, more data is being processed, since increasing the amount of training data will further enhance the performance of the developed deep learning methods