|Ahead of print publication
Radiomics based on biparametric MRI for the detection of significant residual prostate cancer after androgen deprivation therapy: using whole-mount histopathology as reference standard
Zhang-Zhe Chen1,2, Wei-Jie Gu2,3, Bing-Ni Zhou1,2, Wei Liu1,2, Hua-Lei Gan2,4, Yong Zhang5, Liang-Ping Zhou1,2, Xiao-Hang Liu1,2
1 Department of Radiology, Fudan University Shanghai Cancer Center, Shanghai 200032, China
2 Department of Oncology, Shanghai Medical College of Fudan University, Shanghai 200032, China
3 Department of Urology, Fudan University Shanghai Cancer Center, Shanghai 200032, China
4 Department of Pathology, Fudan University Shanghai Cancer Center, Shanghai 200032, China
5 MR Research, GE Healthcare, Shanghai 200032, China
|Date of Submission||28-Sep-2021|
|Date of Acceptance||15-Mar-2022|
|Date of Web Publication||03-May-2022|
Department of Radiology, Fudan University Shanghai Cancer Center; Department of Oncology, Shanghai Medical College of Fudan University, Shanghai 200032
Department of Radiology, Fudan University Shanghai Cancer Center; Department of Oncology, Shanghai Medical College of Fudan University, Shanghai 200032
Source of Support: None, Conflict of Interest: None
We aimed to study radiomics approach based on biparametric magnetic resonance imaging (MRI) for determining significant residual cancer after androgen deprivation therapy (ADT). Ninety-two post-ADT prostate cancer patients underwent MRI before prostatectomy (62 with significant residual disease and 30 with complete response or minimum residual disease [CR/MRD]). Totally, 100 significant residual, 52 CR/MRD lesions, and 70 benign tissues were selected according to pathology. First, 381 radiomics features were extracted from T2-weighted imaging, diffusion-weighted imaging, and apparent diffusion coefficient (ADC) maps. Optimal features were selected using a support vector machine with a recursive feature elimination algorithm (SVM-RFE). Then, ADC values of significant residual, CR/MRD lesions, and benign tissues were compared by one-way analysis of variance. Logistic regression was used to construct models with SVM features to differentiate between each pair of tissues. Third, the efficiencies of ADC value and radiomics models for differentiating the three tissues were assessed by area under receiver operating characteristic curve (AUC). The ADC value (mean ± standard deviation [s.d.]) of significant residual lesions ([1.10 ± 0.02] × 10-3 mm2 s-1) was significantly lower than that of CR/MRD ([1.17 ± 0.02] × 10-3 mm2 s-1), which was significantly lower than that of benign tissues ([1.30 ± 0.02] × 10-3 mm2 s-1; both P < 0.05). The SVM feature models were comparable to ADC value in distinguishing CR/MRD from benign tissue (AUC: 0.766 vs 0.792) and distinguishing residual from benign tissue (AUC: 0.825 vs 0.835) (both P > 0.05), but superior to ADC value in differentiating significant residual from CR/MRD (AUC: 0.748 vs 0.558; P = 0.041). Radiomics approach with biparametric MRI could promote the detection of significant residual prostate cancer after ADT.
Keywords: androgen deprivation therapy; diffusion-weighted imaging; prostate cancer; radiomics
Article in PDF
|How to cite this URL:|
Chen ZZ, Gu WJ, Zhou BN, Liu W, Gan HL, Zhang Y, Zhou LP, Liu XH. Radiomics based on biparametric MRI for the detection of significant residual prostate cancer after androgen deprivation therapy: using whole-mount histopathology as reference standard. Asian J Androl [Epub ahead of print] [cited 2022 May 21]. Available from: https://www.ajandrology.com/preprintarticle.asp?id=344695
Zhang-Zhe Chen, Wei-Jie Gu
These authors contributed equally to this work.
| Introduction|| |
The global incidence of prostate cancer has been increasing in most countries; in addition, 20% of newly diagnosed prostate cancer cases in Northern and Western Europe are advanced or metastatic disease, and in China, this proportion is 68%., Androgen deprivation therapy (ADT) is a key primary treatment for advanced and metastatic prostate cancer and is an important neoadjuvant therapy before radiotherapy and surgery. Currently, the assessment of ADT treatment effect in prostate cancer is mainly based on the serum prostate-specific antigen (PSA) test, but this test has shortcomings. First, it cannot assess the changes in primary and metastatic lesions discriminatively. Second, neuroendocrine differentiation often occurs in hormonally treated prostate cancer, leading to a low PSA even when cancer has progressed, thus nullifying PSA monitoring of such tumors.
Conventional magnetic resonance imaging (MRI) was considered unsuitable for the assessment of prostate cancer after ADT because the signal intensity of prostate tissue on T2-weighted MR imaging (T2WI) is homogeneously reduced, which causes lower contrast between cancerous and normal tissue and overestimation of tumor presence., Fortunately, the last decade has witnessed rapid developments in multiparametric MRI (mpMRI) for the management of prostate cancer, and the Prostate Imaging Reporting and Data System (PI-RADS) based on mpMRI has been proven to have high sensitivity and specificity in prostate cancer detection and localization. Diffusion-weighted imaging (DWI) and T2WI serve as the most important sequences; moreover, the apparent diffusion coefficient (ADC) value itself can further support clinicians in the decision-making process for patients with a PI-RADS score <3 at risk for prostate cancer. The biparametric MRI (bpMRI) protocol based on DWI and T2WI has also shown great promise in the assessment of ADT effects in previous studies.,, However, most of these studies evaluated the effects of ADT by comparing the pre- and post-ADT MRI images without full pathological correlation. The selection and delineation of lesions were based on pre-ADT MRI and focused on significantly visible cancer, leading to bias in the results. Even in a study based on whole-mount pathology, the change in lesion volume between pre- and post-ADT images, rather than lesion appearance on the post-ADT images, was used to assess the treatment effect. Since pre-ADT data might not be available for every patient, the exact efficiency for detecting residual lesions mainly based on post-ADT images needs to be investigated. Moreover, after ADT, T2WI is considered unsuitable for the detection of residual disease in clinical applications, but the change in ADC value in prostate cancer was controversial in previous studies. Some studies have suggested a significant increase in ADC value after ADT,, but one study showed an inconspicuous change. Therefore, a new method is needed to analyze post-ADT images for prostate cancer.
With the rise of radiomics, many studies have investigated the use of radiomics features, especially texture features, for prostate cancer patients; these features have shown a promising ability to differentiate between tumor and benign tissues.,, Texture features for prostate cancer can be derived using gray-level co-occurrence matrices (GLCMs) and aim to separate or classify different tissue types, and different statistical features can be extracted from GLCMs, such as Haralick features. Evaluating prostate cancer MRI data with the GLCM approach might improve tumor detection after ADT. In two recent studies, textural features could distinguish tumors from benign tissues after ADT even in cases with low contrast between the tumor and surrounding tissue,, but these studies were not based on whole-mount pathology and did not assess whether the lesions responded to ADT. The exact value of radiomics methods for detecting prostate cancer after ADT remains unclear.
Therefore, the aim of this study was to use the radiomics method to analyze post-ADT bpMRI images to assess prostate cancer and benign peripheral zone (PZ) tissues, and to investigate the ability of this approach to detect significant residual prostate cancer; in addition, to ensure that all the analyses were based on whole-mount histopathology, only patients who underwent prostatectomy after neoadjuvant ADT were enrolled in the study.
| Patients and Methods|| |
This retrospective study was approved by the Ethics Review Board of Fudan University Shanghai Cancer Center (Shanghai, China; No. 2005217-2). The requirement for informed consent was waived because it was a retrospective study and we used noninvasive methods.
From January 2015 to May 2021, 92 patients who underwent prostatectomy after neoadjuvant ADT and had preoperative MRI scans were retrieved from our history system according to the following criteria: (1) had clinically significant prostate cancer (Gleason score >6, greatest percentage of cancer >50% and more than two positive cores) confirmed by biopsy before ADT, with post-ADT pathology confirmed by radical prostatectomy; (2) underwent pre-ADT bpMRI examinations within 4 weeks before or after biopsy, with bpMRI images of identified prostate cancer; (3) underwent post-ADT bpMRI examinations in our hospital within 2 weeks before surgery; (4) treated for more than 3 months with a complete androgen blockade with bicalutamide plus ADT with goserelin, leuprolide, or abiraterone (based on the discretion of the treating physician); and (5) only underwent ADT.
Among these patients, a total of 60 patients with significant residual lesions confirmed by pathology and 32 patients with pathologic complete response or minimum residual disease (CR/MRD) were enrolled. The clinical data of these patients are listed in [Table 1].
Presurgical MRI was performed with two 3.0-T MRI systems (Magnetom Skyra, Siemens Healthcare, Erlangen, Germany). The entire prostate gland and seminal vesicles were imaged in the axial, sagittal, and coronal planes using a T2-weighted turbo-spin-echo (TSE) sequence (repetition time [TR]/echo time [TE]: 9040 ms/89 ms; number of excitations: 2; slice thickness: 3.5 mm; spacing: 1 mm; and matrix: 320 × 256); T1-weighted imaging was performed with a fast spoiled gradient-echo (FSPGR) sequence (TR/TE: 231 ms/2.46 ms; slice thickness: 5.5 mm; spacing: 1 mm; and matrix: 320 × 240). The DWI images were acquired with a readout-segmented echo-planar imaging (RS-EPI) diffusion-weighted imaging sequence (TR/TE: 6680 ms/66 ms; slice thickness: 3.5 mm; spacing: 1 mm; and matrix: 116 × 116) with identical slice locations to the transverse T2WI imaging; b values were 50 s mm−2 and 1000 s mm−2. The derived ADC map was automatically calculated using a pixel-wise monoexponential analysis of the diffusion-weighted images on a workstation (AW4.6, GE Healthcare, Milwaukee, WI, USA).
After radical retropubic prostatectomy, the intact specimens were inked for laterality and fixed in formalin overnight at room temperature. Care was taken in each case to maintain the orientation of each slice of the prostate so that the same side was routinely cut (i.e., the superior or inferior edges for each prostate cross-section), thus allowing for relatively equal spaces between hematoxylin and eosin (HE) sections. Subsequently, 5-μm tissue sections were cut, mounted on glass slides, and stained with HE.
The lesions were assessed and recorded from whole prostate samples by a dedicated central pathology genitourinary pathologist with more than 16 years of experience in genitourinary pathology. All of the lesions were identified by whole-mount pathology in the positive region confirmed by pre-ADT biopsy. The lesions were divided into two categories, significant residual disease and CR/MRD, because patients with MRD and CR have similar prognoses that are much better than the prognosis of patients with significant residual disease. Pathological CR was defined according to the previous literature based on features such as reduction in gland size with decreased glandular density and increased periglandular density, as well as almost complete degeneration of cancerous cells. MRD was considered when the largest cross-sectional bidimension of the residual lesions was shorter than 5 mm. Significant residual lesions were identified as lesions larger than 5 mm. The outlines of the lesions (residual disease and CR/MRD) and benign PZ tissues were drawn on HE slices for further analysis; if MRD lesions were intermixed with CR tissue, they were delineated together. For each patient, one side of the noncancerous PZ confirmed by pathology was also delineated on all slices as the benign tissue control; if the volume of the PZ was reduced due to involvement of the cancer and difficult to assess, it was excluded.
Region of interest (ROI) delineation
ROIs of significant residual disease, CR/MRD, and benign tissue in the peripheral zone were contoured using ITK-SNAP software, version 3.4.0 (www.itksnap.org), an open-source image processing software program. ROIs were confirmed on all bpMRI maps and contoured together by two radiologists with 5 years (ZZC) and 13 years (XHL) of experience in prostate MR imaging. The ROIs were drawn on the MRI images in ITK-SNAP according to the labeled HE slices under the direction of the pathologist (HLG), as shown in [Figure 1] and [Figure 2]. The location and border of the lesions or benign tissues were identified based on the location of the ejaculatory ducts, the dimensions of the prostate, any identifiable benign prostate hyperplasia (BPH) nodules, and the approximate distance from the base or apex. For each lesion, all positive MRI slices confirmed by pathology were included in the ROIs. If the patients had some suspicious lesions or lesions with unclear borders, the pre-ADT MRI images would be referenced for confirmation. Any differences in measurement were resolved by consensus.
|Figure 1: A 65-year-old man with Gleason 4 + 4 = 8 prostate cancer underwent 7 months of ADT before radical prostatectomy. (a) Pre-ADT T2WI shows a lesion with a homogeneously low signal in the right PZ involving the CG (short arrow). (b) The lesion volume was remarkably reduced on post-ADT T2WI (short arrow), and the signal was heterogeneous and lower than that on benign tissue in the left PZ (long arrow) of the prostate. (c) The lesion (short arrow) showed a slightly higher DWI signal than benign tissue in the left PZ (long arrow) of the prostate. (d) The lesion (short arrow) showed a slightly lower ADC value than benign tissue in the left PZ (long arrow). (e) After prostatectomy, pathology showed a lesion (long arrow) consisting of tumor cells with pyknotic nuclei and degenerated tumor cells with foamy vacuolated cytoplasm caused by androgen deprivation-induced programmed cell death, indicating a complete response in prostate cancer. (f) ROIs were placed on lesions (right) and benign tissue (left) on post-ADT T2WI by referring to the pathology results and pre-ADT images. The ROIs were then copied onto DWI and ADC maps, and radiomics analysis was performed using these ROIs. ADT: androgen deprivation therapy; T2WI: T2-weighted imaging; PZ: peripheral zone; CG: central gland; DWI: diffusion-weighted imaging; ADC: apparent diffusion coefficient; ROI: region of interest.|
Click here to view
|Figure 2: A 68-year-old man with Gleason 5 + 4 = 9 prostate cancer underwent 3 months of ADT before radical prostatectomy. (a) The post-ADT T2WI image showed a diffuse and heterogeneous low signal throughout the prostate, and it was difficult to confirm the existence of lesions. (b) On axial high-b value DWI, the left PZ showed high signal intensity, and the right PZ showed moderate signal intensity. (c) The ADC map showed a heterogeneous, low ADC value in the left PZ lesion and a moderate value in the right PZ. (d) Photomicrography of the histopathology slide reveals significant residual disease on both right (yellow circle) and left (blue circle) sides of the PZ. (e) The residual prostate cancer in the left PZ was still obvious and retained representatively malignant architectures. (f) The right PZ was characterized by a reduction in gland size with decreased glandular density and increased periglandular density, and spare tumor cells could be observed in several locations, showing an infiltrative pattern. ROIs were placed on the lesions on T2WI, DWI, and ADC maps according to c for radiomics analysis. ADT: androgen deprivation therapy; T2WI: T2-weighted imaging; PZ: peripheral zone; DWI: diffusion-weighted imaging; ADC: apparent diffusion coefficient; ROI: region of interest.|
Click here to view
Radiomics analysis was performed using Artificial Intelligence Kit software (Artificial Intelligence Kit version 3.0.0.R, GE Healthcare, Shanghai, China). In total, 381 features were extracted for each ROI, including 39 histogram features, 54 texture features, 119 GLCM features, and 169 run-length matrix (RLM) features. The mean ADC value was also extracted as one of the histogram features from the ROI on the ADC map. Since the intensity range of T2WI images was not always consistent among different patients, before radiomics feature extraction, an image normalization process was used to normalize the gray values of T2WI images with Python programming software (version 3.6, Python Software Foundation; www.python.org). In this process, the image was normalized by centering its intensity at the mean value with standard deviation (s.d.). The gray values of the MRI scans were normalized using the standardization method with a scale of 100. The equation for image normalization was as Inormalize= (I − μ1)/σ1 × 100, where I denotes the original intensity of the T2WI image, Inormalize is the normalized intensity of the T2WI image, μ1 is the mean value of the image intensity, and σ1 is the standard deviation of the image intensity.
Next, ROI data of the significant residual disease and benign tissue in each subgroup were randomly divided into training and validation sets, and radiomics features were extracted for each patient in Artificial Intelligence Kit software. The support vector machine-based recursive feature elimination (SVM-RFE) algorithm was applied to order the features and select the optimal features according to their importance, as described in a previous study.
The selected features were then used to build prediction models (training and validation models) for significant residual cancer with the following classifier methods successively using the software: decision tree, naive Bayes, K-nearest neighbor, logistic regression, support vector machine (SVM), bagging, random forest, extremely randomized trees, AdaBoost, and gradient boosting tree. Receiver operating characteristic (ROC) curve analysis was used to assess the efficiency of each model, and the logistic regression method showed a slightly larger area under the ROC curve (AUC) for the validation models than the other methods. Theoretically, the cost function of logistic regression diverges faster than the other classifier methods, so logistic regression is more sensitive to outliers, which might remedy the relatively lower sensitivity and higher variability of MRI for prostate cancer to some extent. Therefore, logistic regression was selected as the final method for differentiation. A similar process was performed on the significant residual disease and CR/MRD lesion data to construct models for predicting significant residual tissue disease, and on the data of CR/MRD and benign tissue to construct models for predicting a CR/MRD tissue. Logistic regression also showed the highest AUCs and was selected as the final method for differentiation.
All of the statistical analyses were performed with dedicated software (Stata Statistical Software, version 10; Stata Corp LP, College Station, TX, USA), and P < 0.05 was considered statistically significant. Independent t-tests and the Chi-square test were applied to determine significant differences in the patients' clinical characteristics. The ADC values of residual disease, CR/MRD, and benign tissue were compared with a one-way analysis of variance (ANOVA) with Bonferroni's correction, which would require a P value of 0.05/3 = 0.0167 or less to be considered significant; if P < 0.0167, a series of paired t-tests were performed between each pair of tissues in the set. For each pair of tissues with significantly different ADC values, ROC analysis was used to differentiate the two tissues in the validation dataset in the radiomics analysis.
The radiomics signatures were entered into the Stata system for logistic regression analysis. In the training set of significant residual cancer and benign tissue data, univariate logistic regression analysis was performed for each potential predictive factor for residual disease. Next, the features found to be statistically significant in univariate logistic regression analysis were then analyzed with multivariate logistic regression analysis for model construction. Similar processes were performed for the significant residual disease and CR/MRD lesion data to construct the logistic regression model for predicting significant residual tissue, and a model for predicting CR/MRD with the CR/MRD and benign tissue data was also constructed.
ROC analysis was used to evaluate the discriminative ability of the models between significant residual disease and CR/MRD tissues, between CR/MRD and benign tissues, and between significant residual disease and benign tissue, and compare the models with the ADC values. The differentiation efficiencies were compared between radiomics features and ADC values based on the AUC value.
| Results|| |
Comparison of clinical data for patients with significant residual cancer and CR/MRD
No significant differences were found between the significant residual cancer and CR/MRD groups in age, initial PSA, M stage, or Gleason score (GS) (all P > 0.05). The post-ADT PSA was higher for patients with significant residual disease (P = 0.019). The median duration of ADT was longer in the CR/MRD patients than that in residual patients (P = 0.003).
Differentiating among significant residual disease, CR/MRD lesions, and benign tissue using the ADC value
In total, 100 significant residual lesions (mean diameter: 1.8 cm, range: 0.5–3.5 cm), 60 CR/MRD lesions (mean diameter: 1.4 cm, range: 0.5–2.7 cm), and 70 benign tissues were included in the final analysis. The distribution of samples in the training and validation sets is listed in [Table 2]. The ADC value (mean±s.d.) of significant residual lesions ([1.10± 0.02] × 10-3 mm2 s-1) was significantly lower than that of CR/MRD lesions ([1.17± 0.02] × 10-3 mm2 s-1), which was, in turn, significantly lower than that of benign tissue ([1.30± 0.02] × 10-3 mm2 s-1), with P = 0.021 and 0.001, respectively [Figure 1] and [Figure 2]. In the validation set, the AUC of the ADC value was 0.792 for differentiating between CR/MRD and benign lesions, 0.835 for differentiating between significant residual disease and benign lesions, and 0.558 for differentiating between CR/MRD and significant residual lesions [Figure 3].
|Figure 3: (a) The AUC of the prediction model with SVM radiomics features for distinguishing CR/MRD from benign tissue and was similar to that of the ADC value in the validation set. (b) The AUC of the prediction model with SVM radiomics features for differentiating significant residual disease from benign tissue was similar to that of the ADC value in the validation set. (c) The AUC of the prediction model with SVM radiomics features for differentiating significant residual disease from CR/MRD was significantly higher than that of the ADC value in the validation set. The numbers of lesion and tissue samples in the validation set are listed in [Table 2]. ROC: receiver operating characteristic; AUC: area under ROC curve; SVM: support vector machine; CR/MRD: complete response or minimum residual disease; ADC: apparent diffusion coefficient.|
Click here to view
Differentiating among significant residual disease, CR/MRD lesions, and benign tissue using the radiomics method
In the assessment of the significant residual disease and benign tissue radiomics data in the training set, the values of five radiomics features (one GLCM feature and one histogram feature from the ADC map, one GLCM feature from DWI, and two GLCM features from DWI) were positively correlated with the risk of significant residual tissue (all P < 0.05; [Figure 4]). The prediction model for residual lesions based on these features showed an AUC of 0.921 for the detection of significant residual lesions. In the validation set, this model showed an AUC of 0.825 for the detection of significant residual lesions, which was similar to that of the ADC value (P = 0.917; [Figure 3]).
|Figure 4: Distribution and ranking of optimal radiomic features for distinguishing (a) between significant residual and CR/MRD lesions, (b) between significant residual disease and benign tissue, and (c) between CR/MRD and benign tissue. The name of each feature was listed on the left side of corresponding column. Inverse Difference Moment, GLCM energy and GLCM entropy measure the local homogeneity, overall homogeneity and randomness of gray levels of image respectively, in one or more directions. The offset number represents the number of interval pixel between the neighbor point of measurement. Short-run emphasis measures the distribution of the short homogeneous runs (small batch of pixels) in an image, and correlation measures the similarity of the grey levels in neighboring pixels. Those parameters were listed in a format of “name_direction_offset number”. The column represented the contribution of each feature to the differentiation task. The higher the column was, the greater the contribution. The color of each column represented the MRI modality. CR/MRD: complete response or minimum residual disease; MRI: magnetic resonance imaging; DWI: diffusion-weighted imaging; ADC: apparent diffusion coefficient; GLCM: gray-level co-occurrence matrix; s.d.: standard deviation.|
Click here to view
In the assessment of the CR/MRD and benign tissue radiomics data in the training set, the values of two radiomics features (one histogram feature from the ADC map and one GLCM feature from DWI) were positively correlated with CR (both P < 0.05). The values of three radiomic features (one RLM feature and two GLCM features from the ADC map) were negatively correlated with CR (all P < 0.05; [Figure 4]). The prediction model showed an AUC of 0.853 for the detection of CR/MRD. In the validation set, this model showed an AUC of 0.766 for the detection of CR/MRD, which was similar to that of the ADC value (P = 0.672; [Figure 3]).
In the assessment of the CR/MRD and significant residual disease radiomics data in the training set, the values of four radiomics features (two GLCM features from DWI and two GLCM features from the ADC map) were positively correlated with CR (all P < 0.05). The values of five radiomics features (two GLCM features from DWI, two GLCM features from the ADC map, and one RLM feature from DWI) were negatively correlated with the risk of significant residual disease (all P < 0.05; [Figure 4]). The prediction model showed an AUC of 0.854 for the detection of significant residual tissue. In the validation set, this model showed an AUC of 0.748 for the detection of significant residual cancer, which was significantly higher than that of the ADC value (P = 0.041; [Figure 3]).
| Discussion|| |
Our study proved that radiomics methods combined with bpMRI could assess the appearance of prostate cancer after ADT and differentiate significant residual prostate cancer from benign and CR/MRD tissues.
In our study, the ADC values of significant residual disease and CR/MRD cancer were significantly lower than those of benign tissue, similar to previous studies.,, After ADT, the ADC of benign prostate tissue declined due to glandular atrophy, fibrosis, basal cell hyperplasia, and stromal hypercellularity, as well as reduced overall glandular stromal tissue and gland volume. In prostate cancer with a response to ADT, the ADC value increased due to the net decrease in glandular ducts (i.e., net decrease in cellular size or number in tumors) within atrophic prostate cancer tissue as a result of apoptosis. However, the gap between the two tissues remains remarkable. The ADC of significant residual cancer remained low; thus, the differentiation of significant residual disease or CR/MRD tissue from benign tissue was reliable., Such a result could further enable the possibility of delineating cancer in patients using hormonal treatment, which is important for subsequent radiotherapy and other therapies.
The application of radiomics revealed new parameters for the detection of CR/MRD cancer but did not achieve a significantly improved result. On the one hand, this study proved that the ADC value could still play a dominant role in the delineation of prostate cancer due to its reliability and convenience. On the other hand, the radiomics results prove the possibility of using machine learning based on these features to detect cancer in the future, and this approach would be faster than human diagnosis.
In the differentiation of significant residual lesions and CR/MRD tissue, the ADC value showed a low efficiency. The ADC of significant residual cancer was significantly lower than that of CR/MRD, but the gap was minimal, limiting the differentiation of the two tissues. This outcome might be explained by how we drew the ROIs based on the whole amount of pathological data, which included different percentages of benign tissue. Moreover, during ADT therapy, cell death and atrophy of the gland occur simultaneously, and histological changes vary in the initial months of ADT, rendering the change in ADC more uncertain. Thus, the overall change in ADC value might not be sufficient for the assessment of cancer after ADT.
No features from T2WI added to the efficiency of distinguishing significant residual disease from CR/MRD tissue, consistent with the previous view that a change in T2WI signal in prostate cancer after ADT, could nullify the detection of residual lesions. The textural features of the ADC map and DWI were more sensitive to such complex changes. Significant residual tissue is associated with lower short-run emphasis from DWI and inverse difference moments from DWI and ADC maps. Short-run emphasis is employed to measure short-run distribution. The inverse difference moment (IDM) measures local homogeneity and is high when the local gray level is uniform and the inverse GLCM is high. Regions with significant residual tissue always contain more complex tissue, such as glandular tumors with different responses to ADT and atrophy of the gland, thus increasing local heterogeneity and reducing these parameters. Interestingly, in terms of the texture parameter correlation and GLCM entropy offsets, the correlations between pathology and these parameters of DWI were all opposite to those for the ADC map. Image-based correlation measures the similarity of the gray levels in neighboring pixels. Entropy is a measurement of the randomness of intensity images. Although the mechanism of such differences remains unclear, these measures contributed 8 of the 11 effective parameters. These textural features are sensitive to minimal changes and the homogeneity of the whole tissue and thus show more details than the overall ADC map. The results point to the importance of radiomics features on DWI and ADC maps for identifying the status of cancer after ADT. The assessment of prostate cancer and detection of residual disease after ADT are important to the planning and prognosis prediction of subsequent therapy, even in patients who will undergo surgery.
In recent years, prostate-specific membrane antigen-positron emission tomography/computed tomography (PSMA-PET/CT) imaging has been applied to patients treated with ADT to assess treatment response and for the early detection of castration-resistant lesions, and some studies have demonstrated that PSMA-PET/CT might be a suitable quantitative imaging modality for patients after neoadjuvant ADT., However, PSMA-PET also has disadvantages, such as false-negative findings in tumors with no or faint PSMA expression, lower resolution for visualizing the prostate structure, and higher cost than MRI. Therefore, we believe that in clinical practice, radiomics methods combined with bpMRI remain a potential method for assessing the response to ADT in prostate cancer.
To the best of our knowledge, studies of radiomics based on bpMRI for the detection of residual lesions have not been reported before, but there have been a few radiomic studies comparing lesions and normal tissue after ADT. Our study was partly consistent with previous studies in which radiomics could provide exact efficiency for the differentiation of cancer and benign tissue, even after ADT; moreover, the number of textural features included in our study was similar to that in the study by Hedgire et al. but was much smaller than that in the study by Daniel et al. The reason might be because the ROIs were delineated on T2WI with reference to the ADC value after ADT, which might have caused bias because only lesions with remarkable residual disease and dense structures were included, thus increasing the difference between tumors and benign tissues. Meanwhile, in our study, the volume of the PZ shrank after ADT, which reduced the number of voxels; thus, there were fewer effective features associated with the tissues in the PZ.
There were some limitations in our study. First, the study design was retrospective, which might have led to selection bias. For instance, most of the patients who underwent surgery after ADT completely or partly responded to ADT, and patients with progressing diseases were seldom included because of the lack of surgery. The ADT protocol and duration also varied, and the patients with CR/MRD had much longer treatment duration than those with significant residual disease. Second, some small lesions might have been missed on MRI after ADT. Third, several patients had more than one lesion, which might have influenced the uniqueness of individual lesions. Fourth, although our results suggested the potential of radiomics methods for the detection of residual prostate cancer after ADT, no definitive conclusions regarding the use of this method in clinical practice can be reached until a larger number of patients are prospectively evaluated.
| Conclusions|| |
Our study proved that a radiomics method based on bpMRI could differentiate significant residual prostate cancer after ADT from CR/MRD lesions and benign tissue, suggesting a new method for the assessment of prostate cancer after ADT.
| Author Contributions|| |
XHL and LPZ conceived and designed the study. XHL, ZZC, and WJG collected the data. HLG performed the pathological analyses. WL and YZ performed the radiomics analyses of images. BNZ performed the statistical analyses. XHL and ZZC wrote the manuscript with input from all coauthors, and LPZ reviewed and made amendments to the manuscript. All authors read and approved the final manuscript.
| Competing Interests|| |
Yong Zhang, PhD, is a research scientist from MR Research department of GE Healthcare, a US company that manufactures medical facilities, especially diagnostic imaging system; his positional title in the company is leading professional band, he does not hold any share of this company, and he declares no competing interests. The other authors declare no competing interests.
| Acknowledgments|| |
This work was supported by the Clinical Science and Research Fund of Shanghai Municipal Health Commission, Shanghai, China (No. 2020040270).
| References|| |
Teoh J, Hirai HW, Ho J, Chan F, Tsoi K, et al
. Global incidence of prostate cancer in developing and developed countries with changing age structures. PLoS One
2019; 14: e221775.
Zhu Y, Wang HK, Qu YY, Ye DW. Prostate cancer in East Asia: evolving trend over the last decade. Asian J Androl
2015; 17: 48–57.
Li J, Siegel DA, King JB. Stage-specific incidence rates and trends of prostate cancer by age, race, and ethnicity, United States, 2004-2014. Ann Epidemiol
2018; 28: 328–30.
Komura K, Sweeney CJ, Inamoto T, Ibuki N, Azuma H, et al
. Current treatment strategies for advanced prostate cancer. Int J Urol
2018; 25: 220–31.
Saini S. PSA and beyond: alternative prostate cancer biomarkers. Cell Oncol (Dordr)
2016; 39: 97–106.
Puca L, Vlachostergios PJ, Beltran H. Neuroendocrine differentiation in prostate cancer: emerging biology, models, and therapies. Cold Spring Harb Perspect Med
2019; 9: a030593.
Chen M, Hricak H, Kalbhen CL, Kurhanewicz J, Vigneron DB, et al
. Hormonal ablation of prostatic cancer: effects on prostate morphology, tumor detection, and staging by endorectal coil MR imaging. AJR Am J Roentgenol
1996; 166: 1157–63.
Mueller-Lisse UG, Vigneron DB, Hricak H, Swanson MG, Carroll PR, et al
. Localized prostate cancer: effect of hormone deprivation therapy measured by using combined three-dimensional 1H MR spectroscopy and MR imaging: clinicopathologic case-controlled study. Radiology
2001; 221: 380–90.
Pepe P, Garufi A, Priolo GD, Galia A, Fraggetta F, et al
. Is it time to perform only magnetic resonance imaging targeted cores? Our experience with 1,032 men who underwent prostate biopsy. J Urol
2018; 200: 774–8.
Pepe P, D'Urso D, Garufi A, Priolo G, Pennisi M, et al
. Multiparametric MRI apparent diffusion coefficient (ADC) accuracy in diagnosing clinically significant prostate cancer. In Vivo
2017; 31: 415–8.
Kim AY, Kim CK, Park SY, Park BK. Diffusion-weighted imaging to evaluate for changes from androgen deprivation therapy in prostate cancer. AJR Am J Roentgenol
2014; 203: W645–50.
Hedgire S, Kilcoyne A, Tonyushkin A, Mao Y, Uyeda JW, et al
. Effect of androgen deprivation and radiation therapy on MRI fiber tractography in prostate cancer: can we assess treatment response on imaging? Br J Radiol
2019; 92: 20170170.
Hotker AM, Mazaheri Y, Zheng J, Moskowitz CS, Berkowitz J, et al
. Prostate cancer: assessing the effects of androgen-deprivation therapy using quantitative diffusion-weighted and dynamic contrast-enhanced MRI. Eur Radiol
2015; 25: 2665–72.
Karzai F, Walker SM, Wilkinson S, Madan RA, Shih JH, et al
. Sequential prostate magnetic resonance imaging in newly diagnosed high-risk prostate cancer treated with neoadjuvant enzalutamide is predictive of therapeutic response. Clin Cancer Res
2021; 27: 429–37.
Barrett T, Gill AB, Kataoka MY, Priest AN, Joubert I, et al
. DCE and DW MRI in monitoring response to androgen deprivation therapy in patients with prostate cancer: a feasibility study. Magn Reson Med
2012; 67: 778–85.
Khalvati F, Wong A, Haider MA. Automated prostate cancer detection via comprehensive multi-parametric magnetic resonance imaging texture feature models. BMC Med Imaging
2015; 15: 27.
Alvarez-Jimenez C, Barrera C, Munera N, Viswanath SE, Romero E. Differentiating cancerous and non-cancerous prostate tissue using multi-scale texture analysis on MRI. Annu Int Conf IEEE Eng Med Biol Soc
2019; 2019: 2695–8.
Stoyanova R, Takhar M, Tschudi Y, Ford JC, Solorzano G, et al
. Prostate cancer radiomics and the promise of radiogenomics. Transl Cancer Res
2016; 5: 432–47.
Daniel M, Kuess P, Andrzejewski P, Nyholm T, Helbich T, et al
. Impact of androgen deprivation therapy on apparent diffusion coefficient and T2w MRI for histogram and texture analysis with respect to focal radiotherapy of prostate cancer. Strahlenther Onkol
2019; 195: 402–11.
Mckay RR, Montgomery B, Xie W, Zhang Z, Bubley GJ, et al
. Post prostatectomy outcomes of patients with high-risk prostate cancer treated with neoadjuvant androgen blockade. Prostate Cancer Prostatic Dis
2018; 21: 364–72.
Petraki CD, Sfikas CP. Histopathological changes induced by therapies in the benign prostate and prostate adenocarcinoma. Histol Histopathol
2007; 22: 107–18.
Ren Y, Zhang X, Rui W, Pang H, Qiu T, et al
. Noninvasive prediction of IDH1 mutation and ATRX expression loss in low-grade gliomas using multiparametric MR radiomic features. J Magn Reson Imaging
2019; 49: 808–17.
Panebianco V, Villeirs G, Weinreb JC, Turkbey BI, Margolis DJ, et al
. Prostate Magnetic Resonance Imaging for Local Recurrence Reporting (PI-RR): international consensus-based guidelines on multiparametric magnetic resonance imaging for prostate cancer recurrence after radiation therapy and radical prostatectomy. Eur Urol Oncol
2021; 4: 868–76.
Vaz S, Hadaschik B, Gabriel M, Herrmann K, Eiber M, et al
. Influence of androgen deprivation therapy on PSMA expression and PSMA-ligand PET imaging of prostate cancer patients. Eur J Nucl Med Mol Imaging
2020; 47: 9–15.
Onal C, Guler OC, Torun N, Reyhan M, Yapar AF. The effect of androgen deprivation therapy on (68)Ga-PSMA tracer uptake in non-metastatic prostate cancer patients. Eur J Nucl Med Mol Imaging
2020; 47: 632–41.
Sheikhbahaei S, Afshar-Oromieh A, Eiber M, Solnes LB, Javadi MS, et al
. Pearls and pitfalls in clinical interpretation of prostate-specific membrane antigen (PSMA)-targeted PET imaging. Eur J Nucl Med Mol Imaging
2017; 44: 2117–36.
[Figure 1], [Figure 2], [Figure 3], [Figure 4]
[Table 1], [Table 2]