Segmentation
The segmentation of the prostate gland is often a necessary step in clinical settings as well as for further image analysis. Therefore, the possibility of automating whole prostate as well as lesion segmentation is of great interest for the potential time-saving and increased reproducibility. A robust automated segmentation could in turn lead to a fully automated post-processing pipeline.
There are interesting studies in the literature that describe the potentials of deep learning to achieve this goal. For example, Wang et al. [32] compared manual segmentation to a novel deep learning-based one. They were able to demonstrate the feasibility of three-dimensional fully convolutional networks and subsequently validate their findings on a public dataset. Another research group [33] proposed a model based on a propagation deep neuronal network by which data from different levels of complexity were extracted and combined so as to obtain a more trustworthy segmentation of the gland and its boundaries. The propagation deep neuronal network was able to outperform baseline deep neural networks, resulting competitive with the reference standard.
Finally, Alkadi et al. [34] applied a unimodal deep learning-based system using only T2-weighted images for the automated segmentation of both the prostate and PCa lesions. Their algorithm was able to achieve an area under the receiver operating characteristic curve (AUC) of 0.995, comparable to that of other multimodal systems in lesion detection while outperforming other proposed two-dimensional and three-dimensional prostate segmentation approaches.
Cancer detection
There is a high interest in analysing the usefulness of ML-based computer-assisted diagnosis (CAD) software in the field of PCa, as it could improve radiologists’ diagnostic performances and reproducibility. For example, already in 2012, PZ maps created by a decision support system model based on endorectal mpMRI appeared to be a promising tool in correctly localising PZ tumours [35].
More recently, Kwak et al. [36] have demonstrated that it is possible to employ radiomics and ML for the analysis of different tissues and cellular densities in the prostate gland to aid in PCa detection. In order to do so, they developed an MRI-based patient-specific prostate mould in order to ensure correspondence of prostatectomy specimens to the images. A significant difference was found in ML-determined prostatic tissue composition between benign and malignant areas. In 2018, McGarry et al. [37] found that the data derived from ten patients was sufficient to obtain a stable fit for ML MRI detection of increased epithelium and decreased lumen density areas, indicative of high-grade PCa. These authors were therefore able to generate lesion detection maps, also validated against whole-mount histopathological specimens oriented with three-dimensional printed slicing moulds.
Another proposal has been based on volumetric ROI analysis of index lesions on mpMRI [38]. These authors evaluated the usefulness of histogram parameters obtained from T2-weighted, DWI and DCE images in combination with a support vector machine (SVM) ML approach for the improvement of PI-RADSv2 scores, significantly increasing the radiologist’s performance. In a recent study [39], comparing the analysis of ADC radiomics with ML analysis and the performance of mean ADC values alone for differential diagnosis of benign and malignant prostate lesions, the resulting accuracy was similar without a statistically significant difference between approaches.
Ginsburg et al. [40] have suggested that different ML predictive models should be developed for the transition zone and PZ, as lesions and normal prostatic tissue have different imaging characteristics in these zones. In particular, they compared two zone-specific algorithms, trained on different radiomic feature datasets, with a zone-ignorant one. Interestingly, they found that while a significant difference in performance was found between the PZ-specific algorithm and the zone-ignorant one, no differences were found for the transition zone-specific algorithm.
A further use case for ML is the differentiation of stromal benign prostatic hyperplasia from PCa in the transition zone. This diagnosis can be quite challenging, especially with small lesions. Statistical analysis of previously established quantitative features (ADC maps, shape, and image texture) demonstrated a high accuracy in the differentiation of small neoplastic lesions from benign ones using either linear regression and SVM classifiers [41].
Another potential advantage of ML CAD-assisted mpMRI is the improvement of inter-reader agreement. Greer et al. [42] compared index lesion sensitivity, specificity and agreement between eight radiologists from different institutions and with various levels of experience. The readers, after a first qualitative examination, reanalysed the same images using CAD: agreement and sensitivity were higher in the CAD-assisted mpMRI, but specificity worsened. Furthermore, the system was more useful for PZ lesions.
Another CAD system was developed by Ishioka et al. [43], although their proposal was based on a convolutional neural network deep learning algorithm. While the accuracy reported in two validation sets was not very high, with AUC respectively of 0.645 and 0.636, further development could lead to a reproducible, automated CAD system. On the other hand, the research unit lead by Wang [44] has shown better results for a deep learning-based fully automated segmentation when compared to a non-deep learning model. The authors assessed their ability to distinguish PCa from benign pathologies such as prostatitis or prostate benign hyperplasia. Their results showed a better accuracy and reliability of deep learning, with a statistically significantly higher AUC than that of the non-deep learning model (0.84 vs. 0.70). Interestingly, their approach was conducted without any need for segmentation of the training set, potentially easing further evolution of their software with added data.
Assessment of lesion aggressiveness
As PCa is frequently indolent, it is of great importance to assess the aggressiveness of detected lesions, and therefore their clinical significance, for patient management [2]. The standard approach of MRI and subsequent targeted biopsy have improved identification of csPCa foci, but significant disease is still missed, as shown in a recent meta-analysis [45]. Texture features have shown in the past potential as biomarkers of PCa aggressiveness [46, 47].
An interesting study was performed on 56 PCa patients in active surveillance who underwent MRI-guided biopsies [48]. The authors assessed differences between patients who were biopsy- and MRI-negative and patients who were biopsy- and MRI-positive, in radiomic features extracted from T2-weighted and ADC images. They subsequently employed the ten selected parameters to construct models to identify subjects who were biopsy-positive while MRI-negative and those biopsy-negative and MRI-positive. Quadratic discriminant analysis enabled to obtain the best accuracy improving, compared to PI-RADS alone, by 80% for the first classification and 60% in the second. While focusing only on the central gland, Li et al. [49] have shown that an SVM approach, trained on six features extracted from mpMRI exams depicting 152 prostate lesions, was able to consistently predict Gleason score. This is especially important as it is one of the main determinants of PCa clinical significance.
For the task of distinguishing indolent PCas and csPCas, deep learning methods have shown some potential for future applications. Zhong et al. [50] compared both a deep learning and a deep transfer-learning algorithm to the performance of PI-RADSv2. In their validation cohort of 30 patients with 47 lesions, it was comparable to the radiologist’s assessment (AUC 0.73 and 0.71, respectively), while outperforming deep learning alone (AUC 0.69). Similarly, Yuan et al. [51] have shown that transfer-learning, in their case applied to the AlexNet neural network, was able to achieve an overall accuracy of 87% for the prediction of lesion Gleason score.
Finally, a novel approach for the detection of csPCa, as defined in the National Comprehensive Cancer Network guidelines, using radiomics in combination with ML, was recently performed by Varghese et al. [52]. A framework for the robust testing of seven ML algorithms with fivefold cross-validation showed that the best classifier was a quadratic kernel-based SVM with an overall accuracy of 0.92 in the validation cohort.
Local staging and pre-treatment assessment
mpMRI also gives important information for pre-treatment local staging. In this setting, proposals for artificial intelligence applications have been more limited. A recent preliminary report has shown that a radiomics-based Bayesian network achieved a high accuracy (AUC 0.88) in the detection of extraprostatic extension of disease in preoperative MRI, using radical prostatectomy as the reference standard [53]. This finding is promising for further development of ML in PCa local staging.
Some novel applications of ML to prostate MRI have been proposed in the setting of treatment planning. In particular, Sun et al. [3] have developed a new type of focal radiotherapy (bio-focused therapy) that requires prior evaluation of some biological features of the tumour, such as cell density and aggressiveness, which relate to therapy response. These have been assessed by the authors non-invasively, with a voxel-wise approach: patients underwent mpMRI before radical prostatectomy, and imaging data were combined with histopathological information to extract tissue features (cell density). They were then able to create models of prostate tissue cell density, used for the targeting of bio-focused therapy in order to achieve better response and lower toxicity. Shafai-Erfani et al. [54] trained a ML algorithm with paired CT and MRI datasets in order to generate synthetic CT images to be used for patient radiation therapy setup and dose calculation. ML proved capable of producing reliable CT images, comparable to the ground truth for both tasks.
Biochemical recurrence
About 30% of PCa relapse after radical prostatectomy [55]. Also, radiation therapy-resistant PCa or recurrent PCa is not uncommon. The elevation of serum levels of prostate-specific antigen after treatment is currently the most employed biomarker of this condition and is known as ‘biochemical recurrence’ [56]. Very complex altered molecular networks lie behind it and, more extensively, PCa recurrence [55]. ML applied to mpMRI has been proposed as a viable tool for early detection of recurrence or prediction of treatment outcome.
Abdollahi et al. [57] investigated how mpMRI ML models could predict the response to an intensity-modulated radiation therapy. Their findings suggest that both pre- and post-treatment radiomic features can give a reliable prediction of therapeutic success, especially if compared to prostate-specific antigen dosage. A recent study [58] has shown that a SVM could accurately predict biochemical recurrence within 3 years of radical prostatectomy, outperforming a linear regression model based on both MRI and the D’Amico patient risk classification. In fact, ML showed an overall accuracy of 92.2% compared to the 79% achieved by linear regression. The potential of ML for this task was confirmed in 2018 by another study [59] that employed SVM, linear discriminant analysis and random forest on radiomic features extracted from T2-weighted and ADC images.