Improving Prostate Cancer Diagnostic Pathway With A Multi-Parametric Generative Self-Supervised Approach

Abstract

Prostate Cancer (PCa) is a disease characterized by uncontrolled growth of cells within the prostate gland located in the male reproductive system. In 2020 1.4 million new cases and 375 thousand deaths worldwide from PCa were estimated. It is the fourth most commonly diagnosed cancer among both genders globally. Overall, it’s the second most commonly diagnosed cancer and fifth leading cause of cancer death in men. Artificial Intelligence (AI) is a rapid growing domain with a promising potential for healthcare professionals to draw insights from data, and already Machine Learning (ML) are used in discovery and development of medicines, diagnosis, medical imaging and to train and aid the experts judgement. By using Deep Learning (DL) to discover concealed links between Multi-Parametric Magnetic Resonance Imaging (mp-MRI) and PCa to generate a ground truth can potentially strengthen the certainty of the PCa diagnosis by reducing the chance for misdiagnosis due to inter-rater reliability. ML algorithms can assist with analysis of medical data, with the effect being a reduced workload of the skilled workforce. However, many AI algorithms and particularly DL often benefits from large amount of annotated data which can be problematic to acquire in the medical domain due to various reasons such as permission, transparency and privacy. SelfSuperviced Learning (SSL) in computer vision involve performing two tasks, by firstly pretraining a model with unlabeled data with the aim of using the learned latent representation for a different task. Transfer Learning (TL) could be capable to tackle the issue of data scarcity commonly found in the medical domain. This thesis explores the potential of improving the diagnostic pathway of prostate cancer with a multi-parametric generatvie self-supervised approach, by studying whether mp-MRI is effective for classifying clinical significance of PCa. The proposed method uses three-dimensional Apparent Diffusion Coefficient (ADC) and T2-Weighted (T2w) in multi-parametric models with U-shaped Convolutional Neural Network (U-net) architecture and Visual Geometry Group (VGG)16 backbone to explore whether utilizing generative Self-Superviced (SS) approach is useful to classify with a small amount of data compared to a (Fully) Superviced (FS) approach. In this thesis anisotropic are combined with isotropic features in the latent space of the U-net bottleneck. The results from three dimensional reconstruction of masked mp-MRI with generative self-supervised learning showed good results. However, the effectiveness of the proposed methodology were seen to struggle with transferring valuable learned similarity between the source domain and target domain. The final evaluation of the self-supervised approach and fully supervised approach did not yield good enough results quantify to support this aim of the thesis. The multiparametric models were seen to behave more stable in data scarce settings compared to models where a single Magnetic Resonance Imaging (MRI) sequence were used.