1 INTRODUCTION

The main physiological functions of bone are to provide structural support for the body, protect internal organs, store calcium and phosphate, and provide an environment for bone marrow (which contains hematopoietic and mesenchymal stem cells for the production of blood cells and connective tissue, respectively) [1]. Bone is the 3rd most common site of metastasis in the body overall [2]. These bone lesions can appear as osteoblastic/sclerotic (bone-forming), marked by increased intensity on imaging; osteolytic (bone-destroying), marked by decreased intensity on imaging; or mixed appearance. Prostate cancer is the most common nondermatologic cancer in men and has the second highest cancer mortality [3]. Bone is the most common site of metastases from prostate cancer, occurring in more than 80% of patients with advanced prostate cancer [4]. These metastatic lesions typically present with an osteoblastic appearance but can also form osteolytic lesions [4, 5]. Tumor cells preferentially affect the highly vascularized regions of bone, including the vertebrae, pelvis, ribs, and proximal ends of long bones [6] (Figure 1).

The seed and soil model, a well-validated model originally proposed by Stephen Paget in 1889, describes how tumor cells preferentially select specific metastatic environments, with bone being one such site for some cancers [7]. Bone metastases in prostate cancer carry a poor prognosis, with a median survival of approximately 40 months and a 5-year survival rate of 33% [8]. Seminal trials such as the Chemohormonal Therapy Versus Androgen Ablation Randomized Trial for Extensive Disease in Prostate Cancer trial stratify patients into low-volume and high-volume by the number and location of metastases, with high-volume patients showing improved response to chemo-hormonal treatment [9]. Patients with metastatic bone lesions are at particular risk for developing a range of complications known as skeletal related events (SREs), which include bone pain, pathological fractures, hypercalcemia, and spinal cord compression [10]. SREs occur in up to half of advanced prostate cancer patients and greatly impair quality of life, mobility, and functionality [11]. Thus, the early detection, treatment, and monitoring of metastatic bone lesions in prostate cancer cases remains paramount. Until recently, nuclear medicine scan (bone scintigraphy) has been the most commonly used modality for detecting malicious bone lesions. It provides several advantages, including [1] the ability to capture detailed functional information like blood flow and metabolic activity, [2] the ability to scan the entire skeleton, unlike computed tomography (CT) or magnetic resonance imaging (MRI), which are fixed at a specific location, and [3] lower exposure to radiation and cost. Though MRI provides distinct capabilities, including higher sensitivity to soft tissue structures, it is subjected to complicated imaging protocols and cost. On the other hand, with higher spatial resolution, CT is highly sensitive to bone morphology; however, it involves adding to the radiation dose to patients. More recently, positron emission tomography-computer tomography (PET-CT) is the imaging modality of choice for visualizing these bone lesions due to its combination of anatomic and functional imaging capabilities. However, this does not restrain radiation exposure and cost [12]. It is crucial to diagnose malignant bone lesions in the early stage to accurately stage and monitor the disease and assess treatment response. Early detection of these bone lesions can be achieved with multiple paradigms, such as prediction, classification, detection, and segmentation, depending solely on the objective. These tools enable clinicians to make informed decisions using quantitative and qualitative assessments to improve disease detection, progression tracking, and management.

Recent advancements in artificial intelligence (AI) in healthcare have witnessed significant strides, from enhanced diagnostic accuracy to personalized treatment recommendations. AI is generally defined as the ability of computers to replicate human intelligence and behavior. Machine learning (ML) is a subfield of AI that utilizes statistical methods to perform tasks often rooted in pattern recognition, often by minimizing a cost function to reduce error. Deep learning is a subfield of ML which specifically uses deep neural networks as a model architecture. Fueled by data, ML models now excel in detecting complex medical conditions, aiding in early disease identification. Machine learning-based methodologies often have a significant role in the automation process across the paradigms. Popular regression-based techniques such as Elastic Net, Lasso, Ridge, and support vector regressors have been extensively used in regression analysis. Furthermore, traditional techniques such as the support vector machine, decision trees, random forest, and K-nearest neighbor (KNN) were the dominant approaches in classification tasks. In segmentation, we can find the wide usage of contour-based, region-growing, and threshold-based techniques. However, ML methods require domain expertise followed by manual feature extraction, and the performance often remains subpar in addressing complex problems. The advent of neural network-driven architectures overtook ML techniques by automatically extracting pertinent features. In particular, convolutional neural networks (CNNs) have shown immense potential in addressing complex problems.

Additionally, radiomics have been applied to a variety of pathologies to improve lesion analysis [13-15]. Radiomics is a broad term but generally refers to the quantitative extraction of features from radiologic images. These features include both first-order features (e.g., pixel intensity) and higher-order features (e.g., the Gray-Level Cooccurrence Matrix and Gray-Level Run-length Matrix) [16]. Radiomic features can be used as input for ML models, alongside manually annotated image qualities and direct image data, to enhance the performance of these models [17]. While these quantitative tools prove promising in analyzing radiologic images, they are mostly restricted to the research domain and have not yet been translated into clinical work. This review discusses the quantitative approaches used to study osseous metastatic prostate cancer lesions on radiologic imaging, focusing on AI and radiomics methods. Shedding light on these methods reveals underdeveloped research topics and identifies the knowledge gaps and steps needed to translate these strategies and their application to clinical practice.

Given the critical roles of bone in both normal physiology and as a common site of metastasis, particularly in prostate cancer, the necessity for precise detection and monitoring of bone lesions becomes evident. Traditional imaging modalities, while useful, have limitations in sensitivity, specificity, and practical application. This landscape sets the stage for exploring innovative approaches such as AI and radiomics. By leveraging the power of ML and deep learning, these methods can offer enhanced diagnostic accuracy and early detection capabilities. Thus, the research objective of this review is to evaluate the quantitative approaches, particularly AI and radiomics, for studying osseous metastatic prostate cancer lesions. This evaluation aims to highlight how these advanced techniques can address existing limitations in clinical practice, thereby improving disease detection, progression tracking, and patient management.

The inclusion criteria for this review include research papers that cover quantitative methods (radiomics, ML, and deep learning) applied to the study of bone metastatic prostate cancer lesions on radiologic and nuclear imaging (X-ray, CT, MRI, positron emission tomography [PET], PET-CT, and whole bone scintigraphy). The following keyword search was run on PubMed:

(“prostate cancer” OR “prostate metastases” or “prostate lesion” OR ((prostate OR prostatic) AND (cancer OR cancers OR metastases OR metastasis OR neoplasm OR neoplasms OR metastatic OR tumor OR tumors OR malignant OR tumor OR tumors))) AND (radiology OR radiomic OR radiomics OR radiologic OR CT OR MRI OR PET OR PET/CT OR x-ray OR scintigraphy OR imaging) AND (bone OR skeleton OR bones OR osseous OR blastic OR lytic) AND (“machine learning” OR “deep learning” OR “artificial intelligence” OR “neural network” OR “neural networks” OR “radiomic” OR “radiomics” OR “statistical”).

A representative sample of 37 papers covering multiple imaging modalities and quantitative methods were selected for this review. We divide the review by the primary task which the quantitative methods were aiming to achieve. Some of the most common aims of ML include prediction, classification, object detection, and object segmentation (Figure 2a). Some of the most common aims of radiomics include prediction, classification, and heterogeneity analysis (Figure 2b). Figure 3 depicts the number of papers included in the respective methods.

2 MACHINE LEARNING AND DEEP LEARNING TECHNIQUES

2.1 Prediction

Prediction aims to use imaging data to predict future events, such as patient outcomes or treatment response. In early 2008, Chiu et al. [18] presented a perceptron-based neural network for detecting prostate cancer as skeletal and nonskeletal metastasis from technecium 99m-methylene diphosphonate (Tc-99m MDP) whole-body bone scintigraphy images. This study achieved a discriminatory power of area under the receiver operating characteristic curve (AUC) (0.88 ± 0.07, p < 0.001) with sensitivity (87.5%) and specificity (83.3%). However, the inherent limitations of ANNs, such as interpretability and susceptibility to biases, influence the generalizing capability. In a study by Liu et al. [19], the authors experimented with various ML models to predict bone metastasis by analyzing the demographic and clinical data from the Surveillance, Epidemiology and End Results (SEER) database. Out of many methods, the authors discovered that the eXtreme gradient boosting (XGB) yielded good predictive performance. XGB achieved significant results with AUC, accuracy sensitivity and specificity of 96.2%, 88.4%, 90.6%, and 87.9%, respectively. Figure 4a represents the graphical depiction comparing the power of Tc-99m MDP whole body bone scintigraphy images versus XGB in predicting metastases in prostate cancer. Anderson et al. [20] conducted a study investigating patients' survival periods after treating SREs in prostate cancer patients, depicting reasonable performance gain (AUC = 76%) with clinical utility. Constant improvements to these AI-generated, disease-specific prediction models show promise in informing better clinical decision-making in a timelier manner. Namely, increased reliability in forecasting skeletal metastases may allow these prediction techniques to serve as a robust guide for clinicians to formulate personalized treatment strategies derived from predictive outcomes. Consequently, ongoing research endeavors are imperative to enhance the predictive accuracy of ML models, thereby facilitating their seamless integration into clinical practice.

3 Radiomics

Radiomics is an innovative medical imaging field involving the extraction and analysis of quantitative features from radiographic images. Unlike traditional image interpretation, radiomics delves into the intricate details of pixel intensities, textures, and spatial relationships to uncover hidden patterns indicative of underlying physiological or pathological conditions. Zhang et al. [50] proposed a “radiomics nomogram” for predicting bone metastasis from T1-weighted MRI images (dynamic contrast-enhanced). The comprehensive mechanism involved LASSO regression integrated with clinical risk factors, multivariate regression, and decision curve analysis exhibited decent performance (AUC 0.93, CI 0.86–0.89). In a study by Wang et al. [51], the authors analyzed mp-MRI (T2-weighted) and T1-weighted (DCE T1-w) images with various texture characteristics. The authors have used multiple methods (step regression, LASSO regression, and ridge regression) for feature extraction. The authors concluded that combined features from mp-MRI (T2-weighted) and T1-weighted (DCE T1-w) images depicted a better predictive performance with AUC of 0.875 and 0.870, respectively. Reischauer et al. [52] showed how whole-tumor textural analysis changed significantly at 1 month, 2 months, and 3 months of androgen deprivation therapy when compared to pre-treatment measurements and that these changes correlated with serum prostate-specific antigen levels. The authors concluded that employing volumetric texture analysis of the whole tumor could serve as a pertinent approach tackling prostate bone metastasis. Hinzpeter et al. [53] presented a radiomics feature driven by ML algorithms to classify the CT images as metastatic and nonmetastatic using Ga-PSMA PET imaging as a reference standard. Experiments depicted the gradient-boosted tree achieving superior performance with accuracy, sensitivity and specificity of 85%, 78%, and 93%, respectively.

One aspect of prostate bone metastasis radiologic lesion characteristics that has not been explored in depth is inter-patient and intra-patient lesions heterogeneity. A small number of studies have been devoted to prostate cancer bone metastatic lesion heterogeneity on a molecular or histological level, which has shown promise in stratifying lesions by tumor phenotype [54, 55], patient prognosis [56], and treatment planning [57, 58]. Molecular heterogeneity involves multiple factors such as proteins related to androgen receptor sensitivity, DNA repair [54], prostate-specific antigen [55], inflammatory markers, and global DNA methylation. However, we could not identify any studies that focused on radiomic heterogeneity, nor that correlated molecular heterogeneity with its radiomic counterpart. Studies in this area could provide a greater ability to better characterize these lesions and understand the connection between molecular events and appearance on imaging. This facilitates a better understanding of prostate cancer etiology, affecting treatment strategies and perhaps leading to the development of more effective therapies.

4 DISCUSSION

In this discussion section, we critically evaluate the efficacy of ML methods and radiomics in tackling the intricate challenges of prostate bone metastasis detection and analysis. Additionally, we explore avenues such as dataset, heterogeneity, dimensionality, and future research and potential synergies between ML algorithms and radiomic methodologies, aiming to advance diagnostic precision and therapeutic strategies in prostate cancer management.

Machine learning shows immense promise in radiology, possessing the ability to conduct analysis that the human eye is incapable of perceiving. However, there are drawbacks that limit its utility as it currently stands. Machine learning models often struggle with explainability, making it difficult to fully utilize the output or verify its accuracy [50]. Machine learning models also tend to require large amounts of pertinent data for proper training, making generalizability challenging and often requiring each model to only perform a very narrow task. With that said, these limitations will likely be addressed in the coming years with more interpretable methods that can assist radiologists in improving the accuracy and speed of their tasks. Radiomic analysis in prostate bone metastasis offers a noninvasive and quantitative approach, extracting valuable information from imaging data (CT, MRI, etc.). It enables early detection of subtle changes, aids in risk stratification, and provides valuable insights into tumor heterogeneity, allowing personalized treatment strategies. Furthermore, it also aids in predicting treatment response and disease progression. The challenges in radiomic analysis include issues related to the standardization of image acquisition protocols, potential variability in feature extraction methods, and the need for extensive and diverse datasets for robust model training. The interpretability of radiomic features and ensuring clinical relevance remain areas of ongoing research.

With both ML and radiomics, the quality of the output is limited by the quality of the input. Two important characteristics of datasets for both ML and radiomics are the size and dimensionality of the dataset. The size of the dataset is crucial for ML to prevent overfitting and better promote generalizability of the model. For radiomics, large dataset size improves the accuracy and statistical power of extracted radiomic features. Dataset sizes in medical tasks tend to be limited due to Health Insurance Portability and Accountability Act (HIPAA) regulations and minimal public datasets. Data dimensionality plays an important role as well. 3D data contains more spatial information than 2D, but also necessitates higher computational intensity and increased workload if manual annotation is required. Additionally, the quantitative features or model architectures used to analyze 3D versus 2D data differ significantly, making dataset dimensionality an important factor in analysis.

The datasets covered in this review vary depending on the problem-solving approach (prediction, classification, detection, and segmentation) from several 100s of cases to very few. In a predictive analysis study by Liu et al. [19], the authors used a whopping 207,137 cases involving Demographic and clinicopathologic variables extracted from the SEER database to identify metastatic cases. The range (highest and lowest) of patients employed in the classification study falls between 5342 and 21 cases; in detection, 205 to 194; and in segmentation, 205 and 193 cases. However, we did not observe a standard benchmark public dataset across the approaches. Many studies have conducted experiments on in-house and small datasets, challenging the generalizability. Further, we can observe the practice of cross-validation (leave one out, 5-fold, 10-fold) in most of the methods, depicting the stability of the quantitative analysis in the respective experimentation setup [20, 27, 32, 33, 35, 36, 42], and [53]. However, true generalizability can be estimated only by external validation (In Figure 4b, we have shown comparison of sensitivity of AI-based bone lesion segmentation models compared to nuclear medicine physicians). On small data cohorts, we can witness extensive usage of data augmentation techniques by tuning numerous parameters, such as rotation, shift, whitening, and flipping, leading to performance improvement. While augmentation may improve overall performance on the test, it may not enhance generalizing capability. A few methods have also used transfer learning techniques using Imagenet and COCO databases [29, 42]. However, the performance gain through knowledge transfer from nonmedical datasets is still a questionable concern; and coming to various imaging techniques employed for detecting prostate bone metastasis. Bone scintigraphy is the most commonly used modality, while only a few studies have employed MRI, CT, and PET-CT methods. The possible reasons for this are numerous, including [1] the ability to capture detailed functional information like blood flow and metabolic activity, [2] the ability to scan the entire skeleton, unlike CT or MRI, which are fixed at a specific location, and [3] lower exposure to radiation and cost. However, analyzing 3D data is often advantageous as it accurately represents complex anatomical structures and allows volumetric analysis and establishing spatial relationships.

5 CONCLUSION

This review presents multiple strategies, including classification/prediction, detection, segmentation, and radiomic methods for evaluating prostate bone metastasis, helping the community to gain an overview of the literature. We inferred that most methods in the literature achieved reasonable performance (overall accuracy > 0.85), with most of the architectures being benchmark classification and segmentation methods (VGG 16, ResNets, and UNet) with slight modification backed by data augmentation and transfer learning techniques. Our analysis of diverse approaches to addressing prostate bone metastasis provides insight into study frequency in respective disciplines with the performance status. The superior performance depiction in methods without cross-validation and external validation may be subject to overfitting, affecting the generalizability. There are knowledge gaps identified with scope for future study in this field, including lesion heterogeneity analysis, the combination of radiomic and ML analysis for outcome prediction, and detailed lesion characterization for precision medicine.

AUTHOR CONTRIBUTIONS

S. J. Pawan: Conceptualization (equal); data curation (equal); methodology (equal); project administration (equal); writing—original draft (equal); writing—review and editing (equal). Joseph Rich: Conceptualization (equal); data curation (equal); formal analysis (equal); funding acquisition (equal); investigation (equal); methodology (equal); writing—original draft (equal); writing—review and editing (equal). Jonathan Le: Writing—original draft (equal); writing—review and editing (equal). Ethan Yi: Writing—original draft (equal); writing—review and editing (equal). Timothy Triche: Writing—review and editing (equal). Amir Goldkorn: Writing—review and editing (equal). Vinay Duddalwar: Supervision (equal); writing—review and editing (equal).

CONFLICT OF INTEREST STATEMENT

The authors declare that they have no conflicts of interest. V.D. reports consultant activity with Radmetrix, Westat, and DeepTek.

ETHICS STATEMENT

Not applicable.

INFORMED CONSENT

Not applicable.