Nuclear medicine plays a predominant role in the noninvasive assessment of PCa in terms of staging, treatment response assessment, and RLT eligibility assessment, all of which can be improved by radiomics and AI.
Staging
Several authors have attempted to improve the accuracy of nuclear medicine examinations in the staging of PCa patients by applying radiomics, ML, and DL approaches.
In 2019, Zamboglou et al. [53] extracted RFs from [68Ga]Ga-PSMA-11 PET/CT of two cohorts of intermediate/high-risk PCa patients, one prospective (20 patients) and one retrospective (40 patients) cohort, who afterwards underwent radical prostatectomy and pelvic lymph node dissection. RFs extracted from the manual segmentations (GTV-Exp) showed strong correlations with RFs extracted from co-registered histopathological gross tumor volume (GTV-Histo = ground truth; 86% with p > 0.7), both discriminating significantly between PCa and non-PCa tissue. The texture feature QSZHGE discriminated between GS 7 and ≥ 8 for GTV-Exp (prospective cohort AUC = 0.91, validation cohort AUC = 0.84) and also between nodal spread (pN1) and non-nodal spread for GTV-Exp (prospective cohort AUC = 0.87, validation cohort AUC = 0.85). In the multivariate analyses, QSZHGE was a significant predictor (p < 0.01) for PCa patients with GS ≥ 8 tumors and pN1 status.
In 2020, Cuzzocrea et al. [54] performed a radiomics study using 42 high-risk PCa patients staged with [18F]choline PET to assess the relationship between texture analysis of prostatic [18F]choline uptake and patient outcome. For each patient, they calculated the RFs, metabolic parameters of the prostate gland, and the risk assessment score (RAS, based on PSA levels, Gleason score, and T classification). Among 38 RFs, 19 were statistically different between patients with stable disease and patients with biochemical progression at follow-up (p < 0.03). GLCM contrast (Se = 77.8; Sp = 84.8; PPV = 58.3; NPV = 93.3; cutoff = 9.9) and GLZLM-HGZE (Se = 77.8; Sp = 87.9; PPV = 63.6; NPV = 93.5; cutoff = 151.4) showed the best performance for predicting patient outcome (median follow-up 19.8 months), with AUCs of 0.828 and 0.858 (both p < 0.001), respectively.
In 2021, Zamboglou et al. [55] investigated two cohorts of primary PCa patients, a prospective training cohort (n = 20), and an external validation cohort (n = 52). They aimed to find PSMA-PET-derived RFs able to detect intraprostatic lesions missed by visual [68Ga]Ga-PSMA-11 PET/CT assessment. Visual PSMA-PET image interpretation missed 134 PCa lesions (median of 2 missed lesions per patient) with a median maximum diameter of 4 mm (range: 2–6). PCa was missed in 60% of patients in the training cohort (75% with clinically significant PCa, ISUP > 1) and in 50% of patients in the validation cohort (77% with clinically significant PCa). Local binary pattern (LBP) normalized size-zone non-uniformity and LBP small-area emphasis were the only two RFs capable of identifying occult PCa (p < 0.01), with an AUC ≥ 0.93 in the training cohort and AUC ≥ 0.80 in the validation cohort.
In 2015, Gatidis et al. [56] performed the first ML-based radiomic study on 16 PCa patients who underwent staging [18F]choline PET/MRI. A spatially constrained fuzzy c-means algorithm (sFCM) was applied to the single datasets, and the resulting labeled data were used for training a SVM classifier. Accuracy and false-positive/negative rates of the proposed algorithm were determined in comparison with manual tumor delineation or histopathology correlation in 5 of 16 patients. The combined sFCM/SVM algorithm revealed reliable classification results consistent with the histopathological reference standard and comparable to those of manual tumor delineation. Also, sFCM/SVM generally performed better than unsupervised sFCM alone.
In 2021, other authors evaluated the prognostic value of RFs extracted from nuclear medicine images through the help of ML-based techniques. Cysouw et al. [57] conducted a radiomics study on a cohort of 76 intermediate/high-risk PCa patients, who underwent [18F]DCFPyL PET/CT before radical prostatectomy. They aimed to develop a diagnostic ML-based model for detecting the presence of metastases (pelvic lymph node or distant metastases). RFs were chosen via three different feature selection methods: principal component analysis (PCA), recursive feature elimination with random forest, and univariate analysis of variance utilizing the fivefold cross-validation. The resulting random forest algorithm achieved a good discriminatory performance in the detection of lymph node or distant metastasis (both AUC = 0.86, p < 0.01), leading to a noninvasive determination of low-risk patients that could be spared from extended pelvic lymph node dissection.
Papp et al. [58] aimed to investigate the diagnostic performance of dual-tracer ([18F]choline and [68Ga]Ga-PSMA-11) PET/MRI in 52 PCa patients undergoing radical prostatectomy, to predict low-risk versus high-risk lesions (LH) as well as biochemical recurrence risk (BCR) and overall patient risk (OPR) with ML. RFs, extracted from both [68Ga]Ga-PSMA-11 PET and MRI images, in combination with ensemble ML, were applied and compared with conventional PET parameters. The AUC of the ML-based LH model was higher than the SUVmax analysis (0.86 versus 0.80); the accuracies of the BCR model and OPR model were 89% (AUC = 0.90) and 91% (AUC = 0.94), respectively.
Erle et al. [59] aimed to compare and validate supervised ML algorithms to classify pathological uptake in PCa patients based on [68Ga]Ga-PSMA-11 PET/CT images. Authors evaluated 77 RFs from 2452 manually delineated hotspots (1,629 pathological versus 823 physiological, as ground truth) for the training dataset (72 PCa patients) and 331 hotspots (pathological = 128, physiological = 203) for the validation dataset (15 PCa patients). Three ML classifiers were trained and ranked to assess classification performance. A high overall average performance (AUC = 0.98) was achieved, with higher sensitivity for the detection of pathological uptake (sensitivity = 0.97) compared with physiological uptake (sensitivity = 0.82).
The first DL-based study was conducted in 2020 by Hartenstein et al. [60]; they assessed if CNNs can be trained to determine [68Ga]Ga-PSMA-11 PET/CT lymph node status from CT images of 549 PCa patients, evaluating 2,616 lymph nodes identified on PET. The CNN for the binary classification of lymph nodes achieved an accuracy of 89% (AUC = 0.95; Sens = 86%; Spec = 92%) in the training group but failed in the external validation. Hence, this approach is not generalizable, and its value remains unclear.
Moreover, Capobianco et al. [61] developed a DL approach, investigating the use of training information from two radiotracers, [68Ga]Ga-PSMA-11 and [18F]FDG. With limited PSMA-ligand data available, the idea was that the use of training examples from [18F]FDG, a more widely used radiotracer in general oncology, should improve the performance of the DL approach for the assessment of [68Ga]Ga-PSMA-11 images. The CNN network was developed on a larger [18F]FDG PET/CT image dataset (of lymphoma and lung cancer patients), also assessing transfer learning and the ability to encode tracer type. Then, the developed CNN method was trained on [68Ga]Ga-PSMA-11 PET/CT of 173 patients, divided into development (121) and test (52) sets, to both classify sites of increased tracer uptake as non-suspicious/suspicious for cancer and assign an anatomical location. The expert annotations for the N and M status, according to the PROMISE miTNM framework, were used as ground truth. The evaluated algorithm showed good agreement with expert assessment for the identification and anatomical location classification of suspicious uptake in whole-body [68Ga]Ga-PSMA-11 PET/CT.
Finally, Solari et al. [62] evaluated the performance of combined [68Ga]Ga-PSMA-11 PET and mpMRI image biomarker standardization initiative (IBSI)-compliant RFs for the group-wise prediction of postsurgical GS (psGSs) in 101 primary PCa patients, divided into three categories (ISUP grades 1–3, ISUP grade 4, and ISUP grade 5). Nine SVM models were trained: four single-modality radiomics models (PET, T1w, T2w, ADC), three PET+MRI double-modality models, and two baseline models for comparison. A sixfold-stratified cross-validation was performed, and all radiomic models outperformed the baseline models. The overall best-performing model combined PET+ADC radiomics (82%). It significantly outperformed most of the others dual-modality models (PET + T1w: 74%, p = 0.026; PET + T2w: 71%, p = 0.003) and single-modality models, except the ADC-only model (p = 0.138).
Restaging
In 2020, Kang et al. [63] developed a computational methodology using Haralick texture analysis that can be used as an adjunct tool to improve and standardize the interpretation of FACBC PET/CT images to identify BCR, discerning necrotic tissue from radiation therapy and tumor tissue in 28 PCa patients. Four main RFs were chosen and combined with clinical information; the overfitting-corrected AUC and Brier scores of the proposed model were 0.94 (95% CI: 0.81, 1.00) and 0.12 (95% CI: 0.03, 0.23), respectively.
Other authors evaluated different ML-based approaches with different aims. In 2020, Lee et al. [64] examined with an ML-based approach the [18F]fluciclovine PET images of a cohort of 251 PCa patients with suspected BCR following definitive primary therapy, to automatically identify “normal” patients (no disease recurrence) and “abnormal” patients (locoregional or distant recurrence). CNN models were trained using two different architectures, a 2D-CNN (ResNet-50), using single slices (slice-based approach), and the same 2D-CNN with a 3D-CNN (ResNet-14), using a hundred slices per PET image (case-based approach). The best prediction results were achieved by the 2D slice-based CNN (AUC = 0.971, p < 0.001; Sens = 90.7%; Spec = 95.1%). The underperformance of 3D-CNN compared to 2D-CNN could derive from a larger number of learnable parameters in 3D-CNN and would therefore require a larger training dataset size to generate a sufficiently generalizable model.
Moazemi et al. [65] employed five different ML methods on RFs (40 from PET images and 40 from CT images) to classify 2419 [68Ga]Ga-PSMA-11 PET hotspots in 72 patients (48/72 applied for training) as either benign or malignant. Interestingly, RFs assessed in native low-dose CT increased the accuracy significantly. The ML method achieved better accuracy (AUC = 0.98; Sens = 94%, Spec = 89%) than human readers.
Alongi et al. [66, 67] evaluated the potential application of RFs analysis using an ML-based radiomic algorithm to select [18F]choline PET/CT features to predict disease progression in high-risk BCR PCa patients. In their study [67], the authors analyzed 94 high-risk PCa patients who underwent [18F]choline PET/CT restaging imaging to select features able to predict disease progression (median follow-up of 26 months). Discriminant analysis on the RFs extracted yielded an ML model capable of achieving moderate predictive power in the development of nodal (AUC = 69.87, 95% CI 51.34–88.39) or distant metastases (AUC = 74.72, 95% CI 56.36–93.09). HISTO_entropy_log10 and HISTO_entropy_log2 were the two salient features chosen for the discrimination of distant metastases, while GLSZM_SZLGE and HISTO_energy_uniformity were the chosen features to predict nodal metastases.
Bone metastasis
Bone scintigraphy is a reference standard examination to assess bone metastatic spread of PCa patients. In 2021, Cheng et al. [68] aimed to explore efficient ways to early diagnose bone metastasis using bone scintigraphy images through ML methods in two cohorts of 205 PCa patients and 371 breast cancer patients. Authors used bone scintigraphy data from breast cancer patients to pre-train a YOLO v4 with a false-positive reduction strategy and then trained the approach on a dataset of 194 PCa patients under a tenfold cross-validation scheme, which yielded a lesion-level classification sensitivity of 0.72 and a precision of 0.9.
Trying a DL approach, Ntakolia et al. [69] designed a DL method that overcomes the computational burden by using a CNN with a significantly lower number of floating-point operations (FLOPs) and free parameters comparing to other popular and well-known CNN architectures used for medical imaging, such as VGG16, ResNet50, GoogleNet, and MobileNet. The proposed lightweight look-behind fully CNN architecture was used to classify bone scintigraphy images of 778 metastatic PCa patients into three classes: no metastasis, degenerative (defined as the absence of metastasis but presence of degenerative lesions), and metastatic lesions. The final optimal CNN achieved a high accuracy of 91.6% (F1 score = 0.938). Furthermore, the best-performing CNN method was compared to the other abovementioned CNN architectures used for medical imaging, outperforming the others. These results were similar to previous results of the same group [70], using the same CNN architecture for bone scintigraphy images of 586 metastatic PCa patients divided into only two classes (no metastasis and metastasis), resulting in a higher overall accuracy of 97.38% than the one in the previous study.
In this field, also [18F]NaF PET/CT might be useful, having a higher accuracy than bone scan [17, 71]. In 2018, Perk et al. [72] delineated the [18F]NaF PET/CT images of 37 mCRPC patients by an automated algorithm that determines the lesion boundaries based on statistically optimized regional thresholding (SORT). A classification labeled by an expert depending on the likelihood of malignancy (from 0 = background, 1 = definitely benign to 5 = definitely malignant) was applied to 123 bone lesions. Furthermore, the RFs extracted have been used in the ML analysis with nine separate learning methods, where the random forest model performed the best under tenfold cross-validation conditions at discriminating between the 0 + 1 versus 5 class labels (AUC = 0.95, 95% CI 0.93–0.96).
Albeit with a non-bone-specific tracer, Acar et al. [73] using RFs aimed to distinguish lesions imaged via posttreatment [68Ga]Ga-PSMA-11 PET/CT as nonresponding and completely responding (sclerotic lesions) in 75 PCa patients with known bone metastasis. Sclerotic lesions were categorized as complete responding or nonresponding if they showed [68Ga]Ga-PSMA-11 PET uptake levels either below or above liver uptake, respectively. Multiple ML models were developed, and the weighted K-nearest neighbor (KNN) achieved the best classification performance under tenfold cross-validation conditions with AUC = 0.76 (accuracy = 73.5%, sensitivity = 73.5%, specificity = 73.7%).
Finally, in 2021, Hinzpeter et al. [74] investigated the potential application of RFs analysis using an ML-based radiomics algorithm for detecting bone metastases not visible on low-dose CT, extracting from [68Ga]Ga-PSMA-11 PET imaging of 67 patients with PCa as the reference standard (ground truth). The authors analyzed a total of 205 bone metastases with PSMA avidity, but not visible on low-dose CT. The dataset was divided into training, testing, and validation, which allowed the selection of 11 independent RFs. A gradient-boosted tree was trained on the 11 RFs to classify bones as normal or metastatic, using the training dataset. The model achieved a classification accuracy of 0.85 (95% confidence interval [CI]: 0.76–0.92, p < .001) with 78% sensitivity and 93% specificity.
Theragnostics
PSMA-RLT is an emerging treatment modality for advanced PCa [18]. However, almost 30% of patients do not respond to [177Lu]PSMA RLT, which may be due to intralesional and inter-lesional variations of PSMA expression, potentially resulting in undertreatment and reduced RLT efficacy. The early identification of patients who might benefit from RLT can be supported by pre-therapeutic biomarkers derived from radiomics and AI analysis.
In 2018, Khurshid et al. [75] aimed to assess the predictive ability of tumor textural heterogeneity parameters in a total of 328 metastatic lesions from baseline [68Ga]Ga-PSMA-11 PET/CT of 70 mCRPC patients scheduled to undergo [177Lu]PSMA therapy. NGLCM_Entropy showed a negative correlation (rs = -0.327, p = 0.006, AUC = 0.695), and NGLCM_Homogeneity showed a positive correlation (rs = 0.315, p = 0.008, AUC = 0.683) with pre- and post-therapy PSA levels, where a reduction in PSA classified patients as responders (42/70) and an increase in PSA as nonresponders (28/70).
More recently, Moazemi et al. [76] extracted RFs from 2070 malignant hotspots from 83 advanced PCa patients delineated at pre-therapeutic [68Ga]Ga-PSMA-11 PET/CT scan to analyze the OS of patients treated with RLT. Following a LASSO regression feature selection process, the most relevant RFs (PET kurtosis and SUVmin) significantly correlated with OS (r = 0.2765, p = 0.0114).
In 2021, Roll et al. [77] evaluated the predictive and prognostic value of RFs extracted from [68Ga]Ga-PSMA-11 PET/MRI in 21 mCRPC patients before RLT. The PET-positive tumor volume was defined and transferred to whole-body T2-weighted and contrast-enhanced and non-enhanced T1-weighted MRI pulse sequences. Ten independent RFs differentiated well between responders (8/21) and nonresponders’ patients (13/21), and the logistic regression model, including the feature interquartile range fromT2-weighted images, revealed the highest accuracy (AUC = 0.83) for the prediction of biochemical response after RLT. Within the final model, patients with a biochemical response (p = 0.003) and higher T2 interquartile range values in pre-therapeutic imaging (p = 0.038) survived significantly longer.
Finally, Götz et al. [78] investigated how to introduce a dosimetry method where dose voxel kernels (DVK) are predicted by a neural network based on data acquired of the kidneys in 26 patients undergoing therapy with [177Lu]PSMA or [177Lu]DOTATOC, as target organs of the experimental dosimetric method. The method, implemented on SPECT/CT images, was found accurate and competitive when compared to the standard, in which the activity distribution is convolved with a DVK based on a homogeneous soft-tissue kernel.