Virtual Biopsy of Prostate Cancer Using PSMA PET and AI

Prostate cancer is the number one cancer diagnosed in Norwegian men, with a total number of 5118 new cases in 2016. One in seven men will have a Prostate cancer diagnosis by the age of 75. Complications significantly affecting quality of life, such as impotence and incontinence, are common after treatment of prostate cancer even with modern surgical methods and radiation therapy. Focal therapies trying to limit the damage on healthy surrounding tissue is currently being investigated as treatment alternatives even for patients with high-risk prostate cancer. Many malignant lesions in the prostate will however not have clinical significance as they are slow growing.

Grade of malignancy is assessed on histopathological examination by using Gleason grading: A pathologist grades the pattern in the most dominant and second most dominant parts of the prostate cancer. The resulting Gleason Score is the sum of the two grades. Gleason Score, PSA and disease stage evaluated from clinical examination and MRI is used for risk assessment of the malignancy. When deciding between treatment alternatives one must take into consideration the grade of malignancy, and the burden associated with complications after treatment, as well as life expectancy and comorbidity. Although MRI has good sensitivity in finding clinically significant prostate cancer for targeting biopsy, it remains unclear if biopsy can be avoided when MRI is negative. MRI of the prostate is standardized through the PI-RADS V2 recommendations, but image reader experience affects the detection rate of prostate cancer. For patients with low-risk prostate cancer, who are eligible for active surveillance instead of immediate treatment, regular follow-up and repeated biopsies with risk of infection, sepsis and bleeding is necessary, as current imaging does not provide the necessary specificity.

The new PET tracers targeting prostate specific membrane antigen (PSMA) can improve localization of primary tumours as well as suspicious lesions and improve diagnosis of recurrent prostate cancer. PSMA-PET can provide improved assessment of prostate cancer when used in diagnostic imaging, compared to MRI alone. PSMA expressions is increased in prostate cancer cells and correlate with disease aggressiveness. With PET-imaging it is possible to quantify the uptake of radioactive PSMA tracers, even at multiple time points (dynamic PET-imaging). In conventional PSMA PET, a static scan is performed around one hour after tracer injection. By performing a dynamic scan, tracer uptake can be recorded from the time immediately after tracer injection and on succeeding time points.

The PSMA radiotracer that will be used in the study, is one of the most common radiotracers used for detection of prostate cancer. Despite this, recent studies of the biodistribution of this tracer provides limited evidence of its metabolic pathways. Therefore, we intend to perform a more extensive metabolite analysis in this project.

PET scanners are combined with CT or MRI in order to provide anatomical correlation of PET tracer uptake. CT has the benefit of being well suited for use in attenuation correction of the PET images. In contrast MRI is a bit harder to use in attenuation correction, but it has excellent soft tissue contrast, enabling good visualization of the prostate and surrounding soft tissue structures.

Worldwide the availability of PET/CT is high compared to PET/MRI, and PET/MRI is more expensive both in terms of machinery cost, time and competence needed. PSMA PET can be used to direct biopsy and therapy, for example by targeting radiation therapy. PSMA-ligands can also be used in a theranostic approach (combining imaging and treatment) in order to target treatment to cancer cells expressing PSMA with radioactive ligands.

Machine learning algorithms allow computers to learn based on a set of examples. With a methodology known as "deep learning", it is possible to train machine learning classifiers without predefined sets of features - the algorithm finds these features itself as a part of the training process. In recent years deep learning has been a popular tool in a number of applications within medical imaging, also with regards to prostate cancer. There are even online competitions on creating the best machine learning algorithms for analysis of prostate MRI images, such as the Prostatex-challenges for malignancy prediction and the Promise 12 challenge for segmentation. Hybrid PET-imaging combined with machine learning techniques is also an area of increasing research interest, including hybrid imaging with PSMA PET. Both dynamic PET and MRI provide a multitude of data both functional and morphological. All this data is hard to take advantage for nuclear medicine specialist, but it is well suited for a machine learning approaches where an abundance of data is an advantage.

Targeted radiotheranostics (TRT) is a dynamic and rapidly advancing field in cancer treatment that combines the diagnostic power of molecular imaging, primarily by positron emission tomography (PET) or single photon emission computed tomography (SPECT), with the therapeutic capabilities of targeted radioligand therapy (RLT). This innovative approach enables precise targeting and treatment of both localized and disseminated tumours while minimizing damage to healthy tissues. TRT utilizes radioactive isotopes, or radionuclides, which emit radiation suitable for both imaging and/or therapy. These radionuclides can be attached to various of molecules (binders) such as antibodies, peptides, or small molecules, enabling them to specifically target cancer cells or other components of the tumour microenvironment including endothelial cells, fibroblasts or inflammatory cells. The dual functionality of radiotheranostics, using the same compound for diagnosis and treatment, embodies the essence of precision medicine, ensuring that the therapy is delivered directly to the intended target. When administered systemically, TRT can effectively treat metastatic disease. Despite its promising potential, the clinical application of TRT faces several challenges. These include the development of more specific tumour-targeting radiopharmaceuticals, selecting appropriate radionuclides that maximize cancer cell eradication while minimizing effects on healthy tissue, thereby increasing the therapeutic index, and tailoring treatments to individual patients for personalized care. The most common current TRT for prostate cancer utilize [68Ga]- or [18F]-labelled PSMA as the radiotracer, followed by therapeutic treatment with the beta-emitter [177Lu]PSMA. It is well known that internalisation and subsequent long retention time of the therapeutic radioligand is fundamental for successful TRT. In contrast to the reversible surface binding of PSMA to endothelial cells in glioma, kidney or liver cancer, prostate cancer internalises the PSMA ligand via clathrin-mediated endocytosis, providing potential for increased retention time and thus increased therapeutic effect. A common simplified model for PSMA uptake is the irreversible two-tissue compartment model, where the k3 parameter is a surrogate for ligand internalisation. As k3 in the data from these studies vary by a factor of 2-3 within the study population, this suggest either instability in the model fit, or an inter-patient variability of the internalisation rate. Furthermore, due to the limited temporal and spatial resolution of PET imaging systems, the PET-based model for PSMA internalisation is somewhat simplified compared to a more accurate biological model suggested in the literature. Because of the observed variability and limitations with the simplified PET-based model for quantifying PSMA ligand internalisation, ex-vivo validation is necessary before clinical implementation. In this project we aim to establish methodology and validate PET based internalisation with ex-vivo cellular methods. If successfully validated, the internalisation measurements from PET could serve as a non-invasive tool for patient selection for RLT. Patients with high internalisation of PSMA ligand would potentially be better candidates for RLT, compared to patients with low internalisation, because the increased treatment effect of therapeutic agents with increased retention time.

Two different patient groups will be included in the project, as shown in Figure 1:

1. Patients recruited after prostate biopsy (N=120). These participants will be recruited from patients referred to clinical PSMA PET examination at the PET Imaging Center. Candidate participants will be called for an appointment in a written letter, where the study information letter is attached. A copy of the study information will be attached in their appointment letter. Upon their arrival to their appointment, the patients will be able to discuss their participation in the research project with a study technician, where they will also have the opportunity to sign the informed consent. Patients who fulfill the inclusion criteria and who have given their written informed consent to participate in the study, will then undergo the study-specific PSMA PET/MRI examination prior to their the normal clinical PSMA PET examination. The data from both scans will then be included in the research study.

Patients in Group A will be divided into two arms:

1. 15 min dynamic PSMA PET/MRI of the pelvis, followed by 15 min "early" static whole-body PET/MRI, followed by 30 min "standard" static whole-body PET/MRI (N=100)
2. 60 min dynamic PSMA PET/MRI of the pelvis, followed by 30 min "standard" static whole-body PET/MRI (N=20)
3. Patients recruited before prostate biopsy in their ordinary clinical evaluation of Prostate cancer (N=100).

These participants will be recruited from patients referred to the clinical pathway for prostate cancer. The appointment will be held by an urologist, where an examination is conducted, information about the study, and an opportunity to sign the informed consent is given. Patients who fulfill the inclusion criteria and who have given their written informed consent to participate in the study will then be referred to a PET examination before undergoing prostate biopsies. The MRI part of the PET examination will replace the ordinary MRI scan that the patients undergo in their standard clinical evaluation.

The reason for recruitment of Group B is twofold. First, it provides scan data from patients without any previous biopsy. This is important as it needs to be determined whether previous biopsy affects the resulting PET-data and any machine learning algorithms. Second, it widens the spread of patients over the Gleason scale as the patients in Group A only represent high Gleason scores.

The following data will be collected:

1. PSMA PET/MRI imaging data.
2. Written review/description of the medical images by radiologist/nuclear medical specialist.
3. Arterial and/or venous blood sampling during dynamic PET scanning, to facilitate tracer kinetic modelling.
4. Urine samples that will be used together with the blood samples in the metabolite analysis.
5. Clinical data such as PSA blood test values, weight, height and similar variables.
6. Results from histopathological examination after prostate biopsy and/or prostatectomy.
7. Patient registry information from the Cancer Registry of Norway, Norwegian Prescription Database and Norwegian Cause of Death Registry.
8. Fresh tumour tissue will be extracted from UNN Generell kreftforskningsbiobank The data from the Group A, arm 1 (A1) will be used to investigate if an early dynamic PSMA PET scan (0-15min) of the prostate and a following 15 min static whole-body PET scan can replace the standard 60 post tracer injection PET scan in use today, for primary staging and re-staging of prostate cancer patients. The data from Group A, arm 2 (A2) will be used to build a machine-learning-based transformation model of the kinetic parameters from the late dynamic PET sequence from A2 to the early dynamic PET sequence from A1. The data from Group A and Group B will be used in conjunction with results from biopsy and prostatectomy to develop machine-learning-based prediction models of Gleason Score from PSMA PET, thereby facilitating a virtual biopsy model.

The analysis of dynamic PET data (Group A1 and A2) using tracer kinetic modelling requires arterial blood sampling. Therefore, arterial and/or venous blood samples will be collected during dynamic PET scanning. The blood samples collected for the kinetic modelling will, in addition, be used for a metabolite analysis of the tracer. The urine samples will provide additional information of how the tracer is metabolized in vivo. Arterial blood sampling will also allow us to continue to develop the novel approaches from our group for non-invasive, machine-learning-based prediction of the arterial input function using solely image-derived input data.

Fresh tumor tissue will be collected from UNN Generell kreftofoskningsbiobank (REK ref 2012/1198/REK nord) from patients with primary prostate cancer undergoing radical prostatectomy who were previously successfully included in the Virtual Biopsy project and underwent dynamic PET imaging, and who before surgery also gave consent to deliver material to the biobank.