The Feasibility of a Novel Dose Painting Procedure to Treat Prostate Cancer Based on mpMR Images and Hierarchical Clustering


 BackgroundWe aimed to assess the feasibility of a novel dose painting (DP) procedure for treating prostate cancer with dominant intraprostatic lesions (DILs) based on mpMR images and hierarchical clustering with a machine learning technique. MethodsThe mpMR images of 120 patients were used to create hierarchical clustering and draw a dendrogram. Three clusters were selected for performing agglomerative clustering. Then, the DIL acquired from the mpMR images of 20 patients were categorized into three groups to have them treated with a novel DP procedure being composed of three planning target volumes (PTVs) determined as PTV1, PTV2, and PTV3 in treatment plans. The DP procedure was carried out on the patients wherein a total dose of 80, 85, and 91 Gy were delivered to the PTV1, PTV2, and PTV3, respectively. Dosimetric and radiobiologic parameters (TCP & NTCP) of the DP procedure were compared with those of the conventional IMRT and 3DCRT procedures carried out on another group of 20 patients. A post-treatment follow-up was also made four months after the radiotherapy procedures.ResultsAll the dosimetric variables and the NTCPs of the organs at risks revealed no significant difference between the DP and IMRT procedures. Regarding the TCP of three investigated PTVs, significant differences were observed between the DP vs. IMRT and also DP vs. 3DCRT procedures. At post-treatment follow-up, the DIL volumes and ADC values in the DP group differed significantly (p-value<0.001) from those of the IMRT. However, the whole prostate ADC and PSA indicated no significant difference (p-value>0.05) between the DP vs. IMRT. ConclusionsThe results of this comprehensive clinical trial illustrated the feasibility of our novel DP procedure for treating prostate cancer based on mpMR images validated with acquired patients’ dosimetric and radiobiologic assessment and their follow-ups. This study confirms significant potential of the proposed DP procedure as a promising treatment planning to achieve effective dose escalation and treatment for prostate cancer.Trial registrationIRCT20181006041257N1; Iranian Registry of Clinical Trials, Registered: 23 Oct. 2019, https://en.irct.ir/trial/34305

The results of this comprehensive clinical trial illustrated the feasibility of our novel DP procedure for treating prostate cancer based on mpMR images validated with acquired patients' dosimetric and radiobiologic assessment and their follow-ups. This study con rms signi cant potential of the proposed DP procedure as a promising treatment planning to achieve effective dose escalation and treatment for prostate cancer. Background Prostate cancer is not uniformly distributed in the prostate and may involve several areas of the prostate (socalled multifocal) (1). Surgery, chemotherapy, external beam radiotherapy (EBRT), and other adjunctive techniques are applied separately or in combination with each other for treating prostate cancer tumors (2). EBRT is one of the standard techniques used for treating these tumors (3,4). Treatment procedures like intensitymodulated radiotherapy (IMRT) and volumetric modulated arc therapy (VMAT) are known as high exible EBRT methods for delivering dose prescription. With these procedures, the tumors can be exposed to a lethal radiation dose, while organs at risk (OARs) can be appropriately spared. Satisfactory results have been reported for treating low-risk tumors with these procedures with common prescribed doses (5)(6)(7)(8)(9)(10). However, histopathological assessment of prostatectomy specimens reveals intraprostatic lesions (IPLs) also referred to as dominant intraprostatic lesions (DILs). Therefore, the common prescribed dose to the low-risk tumors is not enough to control and treat the DILs requiring much higher dose levels of 80 Gy and more. One of the recognized leading causes of prostate cancer recurrence and consequently radiotherapy failure are the DILs (11,12). Hence, controlling and treating these tumors can be improved by increasing the prescribed dose. By the way, unfortunately, such an approach induces the cost of increased toxicity, since the dose-escalation of whole prostate gland increases the OARs complications such as bladder and rectum and leads to treatment-related toxicity (13)(14)(15). In an overall dose-escalation procedure, a uniform high level dose distribution is used instead of a non-uniform dose distribution known as dose painting (DP) procedure (16)(17)(18). Therefore, dose escalation just to the DILs with the DP procedure can be regarded a logical procedure without signi cant increase in the risk of injury to the OARs.
In DP procedure, based on the features extracted from functional images, the prostate DILs can be de ned for delivering a non-uniform dose boost to improve tumor control without increasing clinical complications (19)(20)(21)(22)(23)(24). To properly control the tumor and prevent more treatment-related toxicity, it is recommended that DILs be identi ed with multi-parametric MRI (mpMR images) including: T2 weighted (T2W), diffusion-weighted MRI (DW-MRI), and dynamic contrast-enhanced MRI (DCE-MRI) sequences. However, the DIL contouring is a manual procedure wherein a radiologist and radiotherapist decide on the DIL areas to have them included in the target based on medical data and mpMR images. For example, an apparent diffusion coe cient (ADC) map and a volume transfer constant (Ktrans) derived from DW-MRI and DCE-MR images have been investigated extensively as prognostic and predictive biomarkers in a wide variety of tumors (25). With such imaging modalities used in this study to distinguish the tumors from normal tissues, the reported sensitivity levels have ranged from 54-84% and 59-87%, while the speci city levels have ranged from 74-100% and 74-84% for the DW-MRI (ADC) and DCE-MRI, respectively. It has also been shown that the use of mpMR images can increase the accuracy of diagnosing prostate cancer and reduce the number of patients who will require repeated biopsies (26). However, mpMR images still renders some limitations. Variability is reported regarding diagnostic accuracy and inter-reader agreement, being generally dependent on reader experience. Therefore, for DIL classi cation, a problem arises on how to pro t from these imaging modalities.
Machine learning algorithms are arti cial intelligence techniques that adapt statistical and probabilistic tools to learn from preceding examples and then predict new trends. Machine learning techniques have already been applied for detecting prostate cancer (27,28). On the other hand, hierarchical clustering is an exploratory statistical method used for identifying groups based on similarity between the acquired data (29). Hierarchical clustering outcomes can be interpreted since its' algorithm is clear, and the relationship between the input and output can be visualized in a dendrogram. This clustering has been extensively studied in many elds, such as functional MRI in connectivity analysis (30). Nevertheless, hierarchical clustering is infrequently applied to mpMR images. Despite the advantages of dose escalation to DILs, the DP method has not yet been generally introduced as a standard method for prostate radiotherapy due to the escalating dose and its' potential complications.
Hence, it appears that multidisciplinary and comprehensive research is needed to utilize clinically DP method and eliminate existing doubts. According to our knowledge, there is no report on the feasibility of DP procedure to treat DILs based on mpMR images and hierarchical clustering with machine learning. Therefore, in this study hierarchical clustering was used for the classi cation of DILs in prostate cancer patients. Then, the DP planning was performed on a group of patients undergoing prostate radiotherapy based on classi cation of their DILs based on mpMR images and hierarchical clustering with a machine learning technique. Finally, to investigate the feasibility of DP procedure, the dosimetric and radiobiologic parameters of the DP treatment plans were extracted and compared with those of another group of 20 patients treated with the conventional IMRT. In addition, for the patients treated with the IMRT, a 3DCRT planning was also investigated. Besides, a post-treatment follow-up was performed four months post-radiotherapy for the two group of the patients to assess their tumor response to the DP and IMRT treatment procedures.

Materials And Methods
This clinical trial study was approved by our institutional ethics committee and registered by national registry of clinical trials. All the procedures carried out in this research were in accordance with the Helsinki Declaration (1964) and its' amendments. Informed consents were obtained for any experimentation with the subjects. The general framework of the study is depicted as a diagram in Figure 1.

Patients' Characteristics for Designing Hierarchical Clustering
The patients with the following inclusion criteria were regarded for training hierarchical clustering: biopsy-proven prostate cancer with localized intermediate or high-risk disease and no evidence of metastatic disease. The exclusion criteria were: previous prostate radiotherapy, prostatectomy, and contraindications to MRI including, cardiac pacemakers, prosthetic valves, and metal implants. Considering these criteria, the mpMR images of 120 patients, namely the ADC and Ktrans, were used to create hierarchical clustering. Characteristics of the patients are presented in Table 1.

Delineation of the DILs
All the MR images were reviewed by a radiologist (AM, with 11 years of experience in interpreting prostate MRI).
For each case, a combined review of axial T2W, ADC, and DCE images was performed. The DILs were manually segmented by drawing regions of interest (ROIs) along the visible tumor margins in each slice of the ADC maps, with a careful reference to the biopsy-proven regions. The ROI-based mean ADC (10 -3 mm 2 /s) was calculated by averaging the ADC values for the entire slices wherein the DILs were visible. The same radiologist manually drew an ROI on the right external iliac artery of the DCE images for every patient. The AIF was obtained from postaveraging of the ROI at each time point. Subsequently, the DILs were segmented manually based on the corresponding ADC map by drawing an ROI around the enhanced DILs in each slice. Then, the standard Tofts model was applied to generate the Ktrans.

Hierarchical Clustering for mpMR images
After calculating the ADC and Ktrans, the DILs were divided into different levels of risk. Hence, for categorizing the DILs, the hierarchical clustering was used. Categorizing data into clusters is the primary purpose of clustering algorithms, such that similar objects are grouped in the same cluster according to speci c metrics. De ning the optimal number of clusters in a data set is critical in partitioning clustering, such as agglomerative clustering, which requires a user to specify the number of clusters, k, to be created. There is no precise response to this question. The optimal number of clusters is anyhow subjective and depends on the technique applied for estimating similarities and the parameters used for partitioning. A famous and straightforward solution involves investigating the dendrogram produced using hierarchical clustering to see if it suggests an appropriate number of clusters. Hierarchical clustering methods start from many clusters that are objects. The objects are joined gradually into the clusters, up to the nal cluster obtained from all the objects. In each stage, one or two objects and one or two clusters are merged. The hierarchy is an outcome of the fact that larger clusters are regularly obtained by merging smaller ones. Thus, hierarchical clustering is used to draw the typical result of the dendrogram. A dendrogram is a visualization in the form of a tree showing the order and distances of merges during the hierarchical clustering. However, the challenging problem is that there is no golden method to pick optimal clusters. Therefore, if we want to argue for a certain number of clusters, what we are interested in is a considerable jump in the dendrogram's distance that would be typical. In general, this can be done simply by counting the number of intersections with vertical lines of the dendrogram to get the number of formed clusters based on the chosen cut-off value of maximum distance (Figure 2 (a)). A cut-off value of nine was selected in our study, as the jump was pretty obvious. With horizontal cut at different levels in the dendrogram, it was seen that three clusters are a good selection for clustering ( Figure 2 (a)). Therefore, we selected three clusters for performing agglomerative clustering (Figure 2 (b)). It should be noted that the ward linkage method was used to draw the dendrogram and clustering. Besides, the Euclidean was applied for the distance metrics. After determining the number of clusters by drawing the dendrogram (K=3), the next move is to perform partitioning clustering. It is also necessary to mention that the ADC and Ktrans values needing neither feature selection nor pre-processing methods were applied as the input to the hierarchical clustering. As shown in Figure 2 (b), the DILs were partitioned into three groups, and the patients' treatment with the DP procedure was started by creating three levels of risk or planning target volume, including: PTV1, PTV2, and PTV3. The PTV3 and PTV1 were regarded as the high-risk and low-risk prostate regions, respectively, while regions with an intermediate DIL probability as the PTV2. Finally, the radiologist and oncologist concluded that PTV3 correlates with ADC <0.8 mm 2 /s and Ktrans> 3.
In addition, the PTV2 and PTV1 correlated with 0.65 <ADC <1 mm 2 /s, 1<Ktrans<3, 1<ADC mm 2 /s, and Ktrans<1, respectively. In summary, classifying the DILs for delivering different levels of treatment doses in the DP procedure was provided by our in-house clustering algorithms as described above.

Treatment of Patients
Patients' Characteristics The treatment was carried out on two groups of patients, including 20 patients treated with the DP procedure and 20 other patients with the conventional IMRT procedure. The patients' characteristics are listed in Table 3. The mpMR images data is displayed for one of the patients in Figure 3, including the T2W, DWI-MRI, ADC map, DCE-MRI, Ktrans, and reconstructed map. Therefore, based on the ADC and Ktrans parameters, the DILs were classi ed and their dose boost were de ned. For all the patients treated with the IMRT procedure, an alternative 3DCRT planning was also investigated. For the treatment of patients, the inclusion and exclusion criteria were regarded similar to those described before for the training hierarchical clustering. , and PTV3, respectively. Implementing the DP procedure was impossible for ve patients because their prescribed dose led to rectum's overdosage. These patients were excluded from the group of 20 patients who underwent the DP procedure. For the IMRT and 3DCRT techniques, the relevant margins chosen for the prostate were 6 mm along the posterior, and 7 mm along the cranial-caudal, transverse, and anterior directions. In addition, a 10 mm margin was used for seminal vesicles. For both the IMRT and 3DCRT procedures, a total dose of 80 Gy (2Gy/fraction) to the whole PTV was planned. For the 3DCRT procedure similar to the IMRT and DP procedures, the seven-eld technique was applied using the same 6 MV photon beam.

Dosimetric and Radiobiologic Evaluation
The plans were evaluated using the dose-volume histograms (DVHs) derived from isodose distributions. Based on the DVHs and the dose constraint presented in Table 4, relevant dosimetric variables were calculated and reported for the PTVs and OARs volumes. For biological evaluation of the plans, BioSuite (Updated 10-01-2018) software was used (33). The NTCP was estimated by using the relative seriality model (34). This model describes the response of an organ with a mixture of serial and parallel structure. The following equation gives the NTCP: where vi is the fractional organ volume receiving a dose Di and P (Di) is the complication probability. Relevant parameters used in the model for bladder, rectum, and femoral heads are shown in Table 5 in which the 50% response dose is named as D50, γ as the maximum normalized value of the dose-response gradient, and s describes the relative contribution of each type of architecture which is equal to unity for a fully serial and zero for a fully parallel organ. In addition, the α/β ratio is a measure of the fractionation sensitivity of the cells and the organ-speci c dose. The TCP was calculated for three PTV using the LQ-Poisson "Marsden" TCP model (35) with the following parameters: α =0.155 Gy -1 , α/β=1.5 Gy, α spread=0.058 Gy -1 , and clonogenic density= 10 7 cm -3 .

Statistical Analysis
Data was analyzed using the GraphPad Prism software (GraphPad, USA). D'Agostino test was used to assess the normality of data. One-way ANOVA and Kruskal-Wallis statistical methods were applied for multiple comparisons. The con dence interval (CI) and mean rank were used as statistically signi cant indexes for oneway ANOVA and Kruskal-Wallis statistical methods. For analyzing and comparing the follow-up data of the preand post-treatment, the paired t-test was used. An independent t-test was also used to compare the response to the treatment resulted from the DP and IMRT procedures. P-values less than 0.05 were considered statistically signi cant for paired t-test and independent t-test. The attributed number to the pre-treatment group was chosen 100. Then, the whole prostate volume, DIL volume, ADC value of the whole prostate, ADC value of the DIL, and PSA for the post-treatment groups were computed as the percentages of it. The vertical bars in histograms represent the standard deviation (SD) of the means.

Results
Dosimetric and Radiobiologic Analysis Table 6 shows the dosimetric variables of the OARs for the DP, conventional IMRT, and 3DCRT procedures. The differences of the OARs' doses among the DP, IMRT, and 3DCRT procedures were signi cant. For all the dosimetric variables, no signi cant difference was found between the DP and IMRT procedures. For the bladder, the mean of dosimetric variables for the DP was slightly higher than the IMRT. But for the rectum, the mean of the dosimetric variables was lower for the DP procedure. Figure 5 shows the cumulative dose-volume histograms (DVH) of the prostate and OARs for the DP, IMRT, and 3DCRT procedures. As shown, the bladder, rectum, and femoral heads dosimetric parameters are improved signi cantly with the DP procedure compared with other procedures. Although there was no signi cant difference between the DP and IMRT for femoral heads, the mean variables were lower for the DP procedure. As shown in Table 7, no signi cant difference was observed in the values of NTCP for the OARs between the DP and IMRT procedures. Similar to the dosimetric variables for the bladder, the NTCP of the OARs for the DP was more than IMRT, while it was less for the rectum. There was a signi cant difference in the OARs' NTCP values between the IMRT vs. 3DCRT and the DP vs. 3DCRT. As shown in Table 7, for the TCP in three investigated PTVs, signi cant differences are observed between the DP vs. IMRT and DP vs.
3DCRT. Patients' Follow-up The DIL volume and ADC value for both pre-and post-treatment groups were signi cantly different (p-value< 0.001). Post-radiotherapy, the DIL volumes and ADC values in the DP group signi cantly differed from IMRT (p-value< 0.001) as presented in Table 8 and Figure 6. Pre-treatment, 34 DILs were diagnosed in the DP group from which 30 DILs were treated entirely post-treatment (Figure 7 (A)) and 4 DILs were not completely disappeared ( Figure 7 (B)). However, the number of DILs post-treatment in the DP was less compared with IMRT group. The percentage change of the ADC of the DILs post-treatment of the DP was greater than IMRT. Pre-treatment, 37 DILs were identi ed in the IMRT group. Although a decrease in the volume of DILs was observed post-treatment, the volume of DILs was quite apparent for 26 cases (Figure 7 (C)) in this group. This nding suggests a poor response to the IMRT treatment. There was a signi cant difference for all of the parameters pre-and posttreatment with either the DP or IMRT including: the whole prostate volume (p-value< 0.001), whole prostate ADC (with a p-value of 0.008 and <0.001 for the DP procedure and IMRT procedures, respectively), and PSA (p-value< 0.001) for both groups ( Figure 6). However, post-treatment, comparing these parameters with the DP and IMRT indicated no signi cant difference between them with a p-value of 0.06, and 0.64 for the whole prostate ADC and PSA, respectively, while a signi cant difference was observed for the whole prostate volume (p-value = 0.01). Ktrans extracted from the DWI and DCE-MRI images could be used to treat prostate cancer using the DP procedure. Furthermore, the feasibility of the DP procedure based on mpMR images as an alternative technique for IMRT and 3DCRT in prostate cancer radiotherapy was investigated and analyzed in our study. The treatment planning was compared according to the dosimetric and radiobiological parameters, including the TCP and NTCP.
In the DP group, in addition to the prostate, the TCP was calculated for the DILs for α/β of 1.5 Gy. Radiobiological parameters were calculated using the LQ Poisson "Marsden" TCP model (35) and relative seriality model.
The feasibility of the DP procedure has also been investigated clinically by Lips et al. (20). Based on their mpMR images and drawing two PTVs, a total dose of 77 Gy in 2.2 Gy/fraction has been prescribed to the whole prostate.  (38), MR images of the prostate were evaluated for the DILs assessment. A total dose of 75.6 Gy (1.8 Gy/fraction) was prescribed to the whole prostate, and a dose of 151.2 Gy (200% of the prescribed dose) in 3.6 Gy/fraction was delivered to the DILs. According to their study, the distance between the lesion and the rectum restricted the ability to plan high-dose radiation to DILs. Moreover, they announced that DIL planning seems possible to treat DILs. In our study, the distance between the lesion and rectum was one of the limiting factors of the prescribed dose. Consequently, for ve patients, implementing the DP procedure was not possible because the prescribed doses lead to an overdosage of the rectum. These patients were excluded from the DP treatment. However, generally it was possible to deliver a dose escalation to the DILs without compromising the dose constraints for the rectum.
As reported, the DIL volumes pre-and post-treatment were signi cantly different in both groups. In the DP group, the ratio of therapeutic volumes was much lower than in the IMRT group, and for most DILs, the therapeutic volume was disappeared post-treatment. It can also be said that ADC is one of the essential parameters for examining the response to the post-radiotherapy treatment. Hence, increasing the ADC post-treatment can be considered a good criterion of DIL response to treatment. However, some of the lesions are resistant to high-dose treatment; therefore, the primary treatment regimen cannot completely eradicate the DILs. Furthermore, the volume and ADC of the whole prostate and PSA were measured pre-and post-treatment, and it was found that these parameters were different in the IMRT and DP groups. However, no signi cant difference was detected between the two groups for the whole prostate ADC and PSA.

Conclusion
We performed a novel and comprehensive clinical trial study and illustrated the feasibility of the DP procedure to treat prostate cancer based on mpMR images validated with the patients' dosimetric and radiobiologic parameters and their follow-ups. Our study con rms signi cant potential of the proposed DP procedure as a more promising treatment planning for an effective and complication free dose escalation in prostate cancer. The general framework of study design  CT planning with contours of DILs, prostate, and rectum DVHs of the DP, IMRT, and 3DCRT procedures for the prostate and OARs Although there was no signi cant difference between the DP and IMRT for femoral heads, the mean variables were lower for the DP procedure. As shown in Table 7, no signi cant difference was observed in the values of NTCP for the OARs between the DP and IMRT procedures. Similar to the dosimetric variables for the bladder, the NTCP of the OARs for the DP was more than IMRT, while it was less for the rectum. There was a signi cant difference in the OARs' NTCP values between the IMRT vs. 3DCRT and the DP vs. 3DCRT. As shown in Table 7, for the TCP in three investigated PTVs, signi cant differences are observed between the DP vs. IMRT and DP vs. 3DCRT. The DIL volume, ADC value of the DIL, whole prostate volume, ADC value of the whole prostate, and PSA for both of the pre-and post-radiotherapy groups. The attributed number to the pre-treatment group was chosen 100.

Abbreviations
Then, the whole prostate volume, DIL volume, ADC value of the whole prostate, ADC value of the DIL, and PSA for the post-treatment groups were computed as percentages of it. Vertical bars represent standard deviation (SD) of the mean. *p<0.05, **p<0.01, ***p<0.001.