Background

The current standard of care for prostate cancer (PCa) diagnosis is magnetic resonance imaging (MRI) followed by prostate biopsy.¹ The Prostate Imaging and Reporting Data System (PIRADS) provides guidance for standardization of prostate MRI acquisition, interpretation, and reporting.² However, there are several limitations to MRI as an imaging modality, PIRADS as a standardization system, and radiologists as experts and human beings. Hence, MRI-based artificial intelligence (AI) frameworks emerge as a promising strategy.³ Herein, the authors propose “P^AIRADS” (Prostate Artificial Intelligence Imaging and Reporting Data System), a system that fully incorporates a quantitative AI evaluation of the prostate MRI with clinical parameters such as PSA, family history, race, and ethnicity.⁴ P^AIRADS’ goal is to enhance the accuracy and efficiency of PCa detection, aiming to overcome MRI, PIRADS, and health care providers’ limitations and pitfalls as well as streamline the PCa diagnostic pathway (Figure 1).

Automated MRI Feature Extraction

Radiomics utilizes statistical, shape-based, or texture-based features for extracting a large number of quantitative fixed features from MRI, some of which can be invisible to the naked eye. Radiomics facilitates the conversion of images into high-dimensional, mineable data, and is increasingly integrated with machine learning models for improved diagnostic, prognostic, and predictive accuracy in various clinical settings. Although radiomics is transparent, explainable, and efficient, its performance is limited compared to data-driven methods, which are optimized for more accurate classification. In radiomics, both regions of interest (ROIs) and features must be predefined by humans.⁴

Deep learning (DL) is a subset of machine learning that processes complex patterns in data. Characterized by its multilayered architecture, DL autonomously extracts features for tasks such as image recognition, using large datasets in its training process and reducing the need for manual feature extraction. For example, convolutional neural networks that mimic human neuronal architecture are commonly used for feature extraction and classification. Despite its power and extensive applications, DL has its limitations, such as large amount of data requirements, computational intensity, and low “explainability” in the decision-making processes. It is criticized for its “black box” nature due to a lack of transparent process behind its feature extraction.⁴

Green learning (GL) is a novel non-DL machine learning approach that incorporates the depth of DL with the transparency of radiomics, offering faster, explainable, and energy-efficient modeling, thus reducing the environmental impact of these technologies.⁴ GL requires less training data, making it well-suited for medical applications with limited data availability. Additionally, the model size and parameters are smaller than DL models. However, as a newer approach, it has not yet been as thoroughly well-tested and validated as other approaches. The performance of GL models lies between that of DL and radiomics models.⁴

MRI-Based AI Framework Applications

Prostate and ROI segmentation

Prostate segmentation involves delineating the prostate gland in a slice-by-slice fashion. However, it is labor-intensive and susceptible to both inter- and intra-operator variability, attributable to significant heterogeneity in the prostate anatomy.⁵ Effective segmentation by AI algorithms can reduce the risk of missing a lesion and substantially expedite the process, improving workflow efficiency. For example, a GL model was recently developed at the University of Southern California’s Institute of Urology, using a dataset of 119 prebiopsy T2-weighted images and 5-fold cross-validation.⁶ This model demonstrated robust performance, reflected by Dice similarity coefficients of 0.85, 0.81, and 0.62, alongside Pearson correlation coefficients for volume estimations of 0.92, 0.93, and 0.63 for the whole prostate, transition zone, and peripheral zone segmentation, respectively (all P < .01).

Prostate cancer detection

Following prostate segmentation, the next step is detecting ROIs suspicious for PCa. AI algorithms, particularly those based on DL, can analyze MR images and identify suspicious areas potentially indicative of PCa, often represented through heatmap predictions (Figure 2). These predictions are typically compared to ground-truth MR images annotated by radiologists to assess the AI algorithm’s accuracy.⁷

PCa classification

After an ROI is detected on MRI, the next step is its classification. AI models are trained on datasets with known outcomes, the ground truth, to predict whether an ROI is benign or malignant. This involves comparing the ROI’s characteristics to known patterns of cancerous and noncancerous lesions. Classifications are usually made either as benign or malignant, radiologically according to the PIRADS, or pathologically using the International Society of Urological Pathology PCa Grade Group scoring system.⁸

Open-Source Datasets

The development and validation of most AI algorithms for detecting, segmenting, or classifying prostate ROIs heavily rely on robust datasets. Open-source datasets like The Cancer Imaging Archive and the PI-CAI challenge datasets are valuable resources for researchers, as they are freely available.⁹ These datasets contain a wealth of anonymized patient scans, which are essential for training, testing, and validating AI models. For instance, the PI-CAI challenge dataset contains over 10,000 prostate biparametric MRI exams available for validating AI models in clinically significant PCa detection and diagnosis.¹⁰ However, concerns arise regarding the demographic representation in these datasets as they are not always transparent and might not represent a diverse population.¹¹ Limitations of these datasets include the lack of standardized MRI acquisition, variable imaging quality, variability in quality of annotations (if any), and the absence of clinical variables reporting in some datasets.⁹ Furthermore, the ground-truth for clinically significant prostate cancer prediction is highly variable between datasets. Some datasets use MRI-informed prostate biopsy pathology, others radical prostatectomy specimens, while others don’t clearly define the reference standard. Variations in how biopsies were sampled (such as systematic, target, saturation, or the use of fusion), as well as variations in the accuracy of tissue sampling, also inherently bias the datasets. Nonetheless, having access to such datasets facilitates the accelerated and widespread development of AI algorithms in prostate cancer imaging.

Current Challenges

Major challenges include variability in MRI equipment and protocols across institutions and the diversity of patient populations. Retrospective single-center studies compose most of the evidence so far, warranting the need for prospective and multicenter studies. Additionally, there is significant heterogeneity in reporting the performance of MRI-based AI algorithms in studies, raising concerns about the transparency, validity, and replicability of the developed models.^11,12 Another challenge is the amount of data needed to adequately train these algorithms. Furthermore, not only the quantity but the quality of data is also important, as low-quality data will most likely result in low-quality results.

Conclusion

Artificial intelligence has the potential to increase PCa diagnostic accuracy, reduce rater variability, and improve workflow efficiency while reducing the burden experienced by physicians. However, addressing challenges, including the need for extensive and high-quality datasets, standardization of MRI protocols, diversity in patient populations, and transparency in study reporting, is crucial. A comprehensive system that incorporates MRI, AI, and clinical parameters, P^AIRADS, is essential.