Dynamic Detection
By Keith Loria
Radiology Today
Vol. 26 No. 4 P. 14
Emerging techniques are enhancing prostate MRI.
Pejman Motarjem, MD, diagnostic radiologist and the vice chair of operations in the department of diagnostic radiology at City of Hope National Medical Center in Duarte, California, notes MRI is superior to other modalities, such as CT or ultrasound, in diagnosis and staging of prostate cancer.
“MRI provides detailed information about the prostate tissue and is highly effective in detecting clinically significant cancers, as well as guiding targeted biopsies after diagnosis,” he says. “Unlike CT, MRI does not use radiation and has been shown to have a low risk of missing aggressive cancers in most patients.”
Additionally, MRI is more sensitive for detecting clinically significant cancers than ultrasound and provides superior soft tissue resolution to CT scans. The primary drawback is that MRI is generally more expensive and time-consuming and sometimes may cause patient discomfort compared with CT due to study acquisition time.
One concern is that approximately 5% to 15% of low-grade prostate carcinomas may be missed on prostate MRI, according to Motarjem. Multiparametric MRI (mpMRI) helps address this concern. mpMRI is a specialized imaging technique used primarily for the diagnosis and management of prostate cancer. It combines several types of MRI sequences for tissue characterization working synergistically to provide detailed anatomical and functional images of the prostate gland and any underlying cancer.
“mpMRI enhances the detection of prostate cancer by combining different imaging sequences to provide a comprehensive assessment of the prostate,” Motarjem says. “The key components of mpMRI include T2-weighted imaging, which provides detailed anatomical information; diffusion-weighted imaging, which measures the diffusion of water molecules within the tissue; and dynamic contrast-enhanced imaging, which evaluates blood flow and can differentiate between benign and malignant lesions. This type of specificity has been a breakthrough in both detection and staging leading to a more precise treatment.”
Vikas Kundra, MD, PhD, chief of oncologic imaging at the University of Maryland Medical Center, notes, after biochemical assays such as prostate-specific antigen and derivative tests, MRI is the first test used for both screening for prostate cancer and local staging of the disease. He explains that mpMRI uses several functional and anatomic MRI sequences for detecting the disease, rather than one particular sequence, to increase both sensitivity and specificity. The findings are categorized using the Prostate Imaging- Reporting and Data System version 2.1 (PI-RADS), a standardized method for interpretation and reporting.
“In schemas for how to read prostate MRI, one particular sequence is usually the primary for detection, and additional sequences add information,” he says. “Which sequence is used varies by the portion of the prostate being evaluated; For example, in the PI-RADS screening schema, the primary sequence used for detecting a peripheral zone tumor is diffusion- weighted imaging with apparent diffusion coefficient reconstructions, and dynamic contrast-enhanced imaging is used to assist in upgrading or downgrading the degree of suspicion for tumor.” Nevertheless, the classification of lesions using PI-RADS does have its limitations.
Emerging Details
A recent study demonstrated that using AI algorithms on prostate MRI scans shows potential to enhance cancer detection and minimize variability among observers. According to the research, a deep learning (DL) model achieves performance comparable to that of an abdominal radiologist in identifying clinically significant prostate cancer on MRI. Naoki Takahashi, MD, a radiologist at the Mayo Clinic in Rochester, Minnesota, and study author, explains the main difference between the DL model and existing AI methods is that the model is trained with only patient-level labels (without location of tumor), while the traditional approach (DL segmentation model) requires annotation of lesions.
“A major drawback of these traditional approaches is that the lesion needs to be annotated by a radiologist or pathologist, which is resource intensive and limits the size of the dataset,” he says. “This applies not only at the time of initial model development but also at the time of model reevaluation and retraining after clinical implementation.”
In the study, Takahashi and his team developed an innovative DL model designed to predict the presence of clinically significant prostate cancer without relying on lesion location data. They evaluated the model’s effectiveness against that of abdominal radiologists in a sizeable cohort of patients who were not previously diagnosed with clinically significant prostate cancer and underwent MRI at various locations within a single academic institution. The researchers utilized a convolutional neural network—an advanced form of AI capable of detecting subtle image patterns that elude human perception—to identify clinically significant prostate cancer from mpMRI scans.
Out of 5,735 examinations conducted on 5,215 patients, 1,514 examinations revealed clinically significant prostate cancer. When tested on both an internal set of 400 exams and an external set of 204 exams, the model’s ability to detect clinically significant prostate cancer was comparable to that of seasoned abdominal radiologists. Furthermore, a combination of the DL model’s predictions and the radiologist’s insights surpassed the performance of radiologists working alone in both the internal and external test sets.
The drawback of the proposed method, Takahashi notes, is a lack of accurate lesion localization, although he says the heat map is reasonably accurate, and probably requires a larger number of cases to train the model. In addition, Takahashi says subjective evaluation (PI-RADS) is prone to intra- and interobserver variability, while DL output is deterministic.
“We look to share the model’s outputs with radiologists to evaluate how they integrate these results into their interpretations,” he says. “Additionally, we will compare the collaborative performance of the radiologists and the model against that of the radiologists working independently in predicting clinically significant prostate cancer.”
Improved Evaluation
Due to the precise nature of mpMRI, the trend worldwide is for “MRI first,” meaning the MRI is performed before deciding on a biopsy. This approach can help guide treatment and avoid unnecessary procedures.
“Generally, an mpMRI is recommended for a patient with elevated serum prostate- specific antigen levels, patients with a negative biopsy but where there is still a clinical suspicion for prostate cancer, and those with an abnormal digital rectal exam,” Motarjem says. “In patients for whom surveillance of potential cancer risk is recommended, mpMRI can help monitor disease progression over time and indicate when intervention is needed.”
When evaluating, physicians look for features that indicate changes caused by the disease. MRI can detect whether the cancer has spread beyond the prostate (extraprostatic extension) or into nearby structures such as the seminal vesicles. “On T2-weighted imaging, signs of progression include irregular shapes or bulging of the prostate capsule and asymmetry in surrounding tissues,” Motarjem says. “Diffusion- weighted imaging can show areas where water movement is restricted, which often indicates cancerous tissue. Dynamic contrast-enhanced MRI highlights areas with abnormal blood flow that could signal tumor growth.
“For people in treatment, MRI can track changes in the size of the tumor and whether treatments are helping to shrink it. These MRI findings help us determine whether treatment is effective and guide future care decisions,” he says.
AI’s Role
Throughout the industry, AI is proving helpful for several aspects of prostate cancer detection and treatment. Motarjem says AI models can enhance radiologists’ ability to detect and characterize prostate cancer on MRI images, potentially reducing variability in interpretation.
“AI has been shown to excel in precise tumor volume measurements, which assess both metastasis risk and treatment outcomes,” he says. “AI can help reduce reader variation in prostate MRI interpretation. Also, AI can assist in guiding treatment decisions, such as radiotherapy planning.”
However, he warns AI is meant to support, not substitute, the expertise of the clinician. “Cancers are highly variable, and an experienced specialist understands the importance of personalization in both diagnosis and treatment,” Motarjem says.
Even though the AI model from the Mayo Clinic was tested using an external test set, Takahashi notes it is unclear how well the model performs at an external site when clinically implemented. “The model needs to be reevaluated and possibly finetuned before implanting at an external site,” he says. “To address the lack of accurate lesion localization, we are developing a DL model that combines the traditional method (segmentation method) and our proposed method. We will use a small number of lesion-annotated cases to train the segmentation model and use a larger number of cases without lesion annotation to train the classification model.”
There are also limitations and challenges associated with using MRI for prostate cancer detection and management. The first is cost. MRI scans cost several thousand dollars in the United States, which could be a barrier to widespread adoption. Additionally, not all health care providers have the expertise or equipment for accurate prostate evaluation using MRIs.
“Interestingly, one study found that MRI evaluation was less effective in younger patients, which could lead to misdiagnosis,” Motarjem says. “There are several factors that could cause this anomaly. The peripheral zone of healthy, younger patients exhibits significantly reduced signal intensity on T2-weighted imaging and lower values for diffusion. Moreover, image quality may be significantly degraded secondary to susceptibility artifacts from metallic prosthesis within the pelvis, especially involving the hip joints. Patients with certain pacemakers or AICDs [automatic implantable cardioverter- defibrillators] may not qualify to undergo prostate MRI.”
Kundra notes other challenges include image degradation due to motion by the patient or due to bowel motility affecting the prostate during the image acquisition, artifacts caused primarily by metal, such as in some patients with hip replacements, and hemorrhage due to recent prostate biopsy. Thus, it is often suggested to wait at least five weeks postbiopsy before MR.
“In addition, for some patients, claustrophobia can be an issue and antianxiety medication may be helpful,” he says. “There is also a limit of spatial resolution for all imaging tests, however, at least for screening, since most define clinically significant prostate cancer as greater than 0.5 cm, this is not limiting for lesions in the prostate. However, for disease elsewhere, spatial resolution may become limiting such as in lymph nodes or other sites.”
Advancements Ahead
One recent development in the field is the recently, FDA-cleared functionality for the previously available AI-Rad Companion Prostate MR for Biopsy Support from Siemens Healthineers, which automatically segments the prostate gland in MRI images, marking its outer contour. The software is compatible with various MRI scanners and is designed to integrate into existing clinical workflows. Katharina Schmidler Burk, product manager for the digital and automation business at the company, notes this software allows radiologists to quickly identify and annotate suspect areas for targeted biopsies. MRI-supported procedures can help urologists detect significant prostate cancers, ultimately improving patient care.
“Typically, the urologist has to manually segment the prostate gland and look at the lesions and where they are, and our software, before the new FDA clearance, could automatically segment the prostate gland,” she says. “Now, with the new clearance, we have a new functionality that also segments and quantifies the classification of lesions based on the PI-RADS guidelines.”
A manual workflow is now automized and postprocessed automatically. Physicians can obtain the results from the system and edit those results, if needed. “We can now make life easier for the physicians, as many manual steps are now redundant for them, so there will be efficiency gains,” Schmidler Burk says.
While there is significant potential for the use of AI, the industry is also exploring other specialized fields, such as radiomics, focusing on quantifying and studying subtle variations in medical images. “Advances in radiomics can help with the detection of cancer aggressiveness and personalized treatment planning,” Motarjem says. “Techniques like PET/MRI, combining metabolic and anatomical imaging, may also offer improved diagnostic accuracy.”
Kundra believes prospective MRI-relevant technologies or techniques that may improve the detection and management of prostate cancer in the future include faster image acquisition sequences; improved software for reconstruction such as machine/DL-based image reconstruction, enabling improved resolution, including with more sparse data to improve the speed of acquisition and quality of image reconstruction; automatic motion correction algorithms; and AI-based tools that point to possibly suspicious lesions to assist radiologists with detection of clinically significant prostate cancer. Efforts to standardize MRI protocols and interpretations, along with cost reductions, will also be important for the widespread adoption and effective use of MRI in prostate cancer management.
“The bottom line is that MRI is now regarded as an essential tool for prostate cancer detection and has become a standard of care at all centers where there is an appropriate level of expertise for performing the study as well as interpreting images,” Motarjem says.
— Keith Loria is a freelance writer based in Oakton, Virginia. He is a frequent contributor to Radiology Today.
