Accuracy of tumor segmentation from multi-parametric prostate MRI and ¹⁸F-choline PET/CT for focal prostate cancer therapy applications

The rationale for focal therapy of prostate cancer is based on the concept of minimizing damage to off-target tissues while achieving successful local treatment response. Focal treatment methods include cryotherapy, laser ablation, high-frequency focused ultrasound (HIFU), irreversible electroporation, and focal brachytherapy [1‐4]. The benefit compared to radical treatment is reduced invasiveness that lessens patient morbidity. The key to the success of focal therapies is appropriate patient selection based on an unequivocal, usually solitary, treatment target [5].

For focal therapies, it is imperative not only to accurately differentiate significant from indolent prostate cancer [6, 7], but also to precisely identify the tumor borders. In recent years, multi-parametric prostate MRI (mpMRI) has emerged as a useful tool for the detection and risk stratification of primary prostate cancer and recently has been incorporated into American Urological Association (AUA) algorithms for patients with suspected or low-risk prostate cancer [8‐10]. When used optimally, mpMRI can cost-effectively identify a significantly greater fraction of clinically significant cancers (Gleason score ≥ 7 [≥ 3 + 4]) compared to standard biopsy alone while minimizing the detection of low-risk cancer [11‐13]. These outcomes are important because accurate risk stratification of primary prostate cancer historically has been fraught with overdiagnosis and overtreatment [14‐16]. However, mpMRI is not perfect. It suffers from a large false-positive rate [17, 18], only moderate inter-rater agreement [19], and a steep learning curve [20], each resulting in unnecessary biopsies that drive complication rates and unwanted detection of low-risk disease.

It has recently been demonstrated that the addition of positron emission tomography (PET) with ¹⁸F-choline to mpMRI (PET/MRI) can significantly improve the identification of Gleason ≥ 3+4 prostate cancer over mpMRI alone [21], mainly by improving on the comparatively low specificity of mpMRI [17]. The purpose of this study was to determine the precision of tumor boundary detection (segmentation) by visual inspection of mpMRI and by semi-automatic segmentation based on ¹⁸F-choline PET in patients undergoing prostatectomy. Due to prostate deformations, differing slice thicknesses between imaging and histology, and differing spatial orientation between imaging and whole-mount pathology, registration of medical imaging to pathology from prostatectomy specimen is challenging [22]. To overcome these issues, we first accurately registered in vivo imaging to the prostatectomy specimen using a multi-step approach presented in this paper.

Patient characteristics are summarized in Table 1. Histological tumor volumes were small with a median of 1.37 mL (range 0.15 to 6.31 mL). The Gleason ≥ 3+4 tumor with the lowest volume (case 5 in Table 1) was not identified by both readers and excluded from further analysis. The median time between mpMRI and PET, and the median time between the second MRI scan and the biopsy procedure were both 22 days. The median time from biopsy to prostatectomy was 102 days (range 55–169 days). The histological results from standard biopsies were upgraded on targeted biopsies in 8 of 10 cases. In one case, the targeted biopsy missed a Gleason 3+4 index cancer (case 2) identified on standard biopsy. No significant intratumoral variations in Gleason pattern within Gleason ≥ 3+4 cancers were observed.

Figure 3 demonstrates the accuracy of BPH nodule registrations obtained from pathology, ex vivo MRI, and in vivo MRI. Overall registration accuracy between in vivo MRI and histology was excellent with maximum HD below 2.7 mm and mean MDA below 0.5 mm with limited inter-individual variability as indicated by their respective standard deviations (Table 2). The overlap coefficients Dice and Jaccard were high between in vivo MRI and final pathology. Note that the final registration distance errors (in HD and MDA) between in vivo MRI and histology were less than the sum of their respective uncertainties for each registration step (Table 2).

The results of tumor border segmentations relative to the histological reference standard are listed in Table 3. Visual tumor border segmentation based on MRI and PET-based thresholding substantially underestimated the true tumor volumes (seg1 by 79%, seg2 by 80%, seg3 by 58%). As seen from Fig. 4, regardless of the segmentation method, the true tumor volume has no major impact on the level of underestimation. Also, the underestimated tumor volume of the single high-risk (Gleason 4+5) cancer found (patient 9) is among the range of otherwise intermediate risk cancers (Table 1). Imaging findings of a typical example case are shown in Fig. 5. Similarity coefficients obtained from MRI and PET contours reveal misplaced tumor borders compared to histology (Table 3). Applying an ellipsoid formula to estimate the VOI volume resulted in a significant overestimation of the segmented volume by an average 48, 46, and 29% for seg1–3, respectively (p < 0.004 for each) (Table 3).

As indicated by Dice coefficients ≤ 0.41 and Jaccard ≤ 0.27, agreement between human readers and human vs. PET-based segmentations was low (Table 4). In fact, human readers and PET-based segmentation often identified different regions of the histological tumor volume, which lead to an improvement of the segmentation accuracy when all segmentations were combined (seg4). The underestimated tumor volume of the combined segmentation was significantly smaller than that of the individual human readers (Table 3).

As expected, including a safety margin around the segmentation volumes significantly decreased the underestimated tumor volume (Table 5). The combined segmentation volume (seg4) with a 5-mm safety margin covered the entire tumor volume in five of nine cases; in the remaining four cases, a missed tumor volume of 0.03 ± 0.05 mL (2.04 ± 2.84% of the total tumor volume) resulted (Fig. 5e–h). While the percent of the underestimated volume was small, a contour expansion up to 9 mm was required to cover the entire tumor of irregularly shaped lesions. In contrast, with even larger target volumes at the 10-mm expansion level, complete tumor coverage could not be achieved for all tumors in the study by either human reader or the PET-based routine alone.

The goal of focal prostate cancer therapies is to destroy the entire tumor lesion using a non-invasive, well-tolerated treatment. Such a treatment would preserve normal genitourinary function while providing sufficient therapeutic efficacy. In order to be effective, focal prostate cancer therapies require accurate localization of the disease. The high anatomic detail provided by mpMRI at 3 T appears well suited to provide necessary guidance. However, it has already been recognized that lesion extension is typically underestimated by mpMRI [32]. Therefore, all focal treatment approaches are performed with a safety margin [33]. The key question is how to optimally define such safety margins.

We developed and applied a methodology for objective retrospective registration of whole-mount pathology to in vivo MRI and PET imaging. This method enabled us to determine whether in vivo imaging correctly identified and classified all tumor lesions found at final pathology from prostatectomy specimen [21]. A similar approach was used to determine boundary errors based on visual image inspection on MRI as well as mathematical thresholding techniques on ¹¹C-choline PET [34]. To our knowledge, this is the first report on the accuracy of prostate cancer segmentations from mpMRI and ¹⁸F-choline PET using registered whole-mount histology as the reference standard. For this performance test, we simulated the clinical situation of a radiologist tasked to delineate tumor borders prior to a hypothetical planned ablation, in which readers were aware of the interpretation of the mpMRI and biopsy-based histology but had no knowledge of the ¹⁸F-choline PET results, ex vivo MRIs, or the pathology from the prostatectomy specimen.

Determining the accuracy of prostate cancer segmentation in MRI is difficult. Typically, error metrics are derived from measured inter-observer variability [35]. This is subject to bias because it lacks a reference standard. We took a different approach by registering imaging to histology, thereby enabling error measurements of tumor border delineation. Our results indicated a substantial difference between “perceived” (i.e., by imaging) and “true” (i.e., by histology) tumor borders. Human expert prostate MRI readers and the PET-based routine were equally unsatisfactory. Depending on tumor shape and the identification of the lesions’ center by human readers, minimum contour expansions between 11 and 15 mm would have been needed to cover all Gleason ≥ 3+4 disease. Currently applied standard safety margins are typically 10 mm or less, which our results imply would often be insufficient.

Ouzzane et al. recently noted that only few studies with valid methodology attempted to address the crucial question of safety margins [4]. Depending on the planned focal treatment method and the lesion location within the prostate, 4–10 mm safety margins have been proposed [1, 36, 37]. For example, Ting et al. used irreversible electroporation to treat 25 men with low- to intermediate-risk prostate cancer with a 5-mm minimum safety margin. Short-term follow-up indicated a 24% recurrence rate at 8 months and almost all recurrences were adjacent to the treatment zone [2], suggesting that treatment of larger tumor volumes would have been needed. Safety margins of 4–6 mm near the sphincter muscle have been advocated to avoid incontinence. Ahmed et al. performed a prospective HIFU trial involving 56 men with low- (12%), intermediate- (84%), and high-risk (4%) prostate cancer. Fifty-two men received a 6-month post-treatment biopsy, and clinically significant cancer was found in 10 cases [38]. Other groups advocate hemiablation to achieve better long-term disease control with low morbidity [39]. Based on the available literature, it remains debatable whether focal prostate cancer treatment is a suitable alternative to established whole-gland treatment approaches.

While our registration method was designed to optimize the accuracy of registration between in vivo MRI and pathology, it is not without error. The quantification of registration errors is difficult because the processing of human prostate tissues has to follow established methodological standards. Since the bladder surface and the prostatic apex are undergoing separate histological assessments to evaluate for possible extracapsular extension of cancer, they are not part of the final prostate specimen. Fiducial markers could not be introduced into the specimen after prostatectomy because they would be dislodged during sectioning of the whole-mount blocks. Due to the stability constraints of plastic molds, slits for gross sectioning of the prostate within the mold were spaced at 3-mm intervals. Therefore, there was a maximum uncertainty of 3 mm and a mean uncertainty of 1.5 mm with respect to the presence or absence of a given lesion on consecutive slices. Using software interpolation between registered whole-mount HE slices, as done in this study, will however lessen uncertainties in tumor border definition. Deformations of the specimen relative to the in vivo geometry of the prostate are often non-uniform. Bending and warping energies are particularly high at the lateral peripheral zones of the prostate [22], which increases registration uncertainties at these locations. While registration errors of each registration step may add up, it is unlikely that they are always to the disadvantage of the lesion representation on in vivo T2W MRI used for tumor segmentation. In fact, when determining the registration errors of BPH nodules from in vivo over ex vivo MRI to histology, the HD and MDA of the final registration (in vivo MRI to histology) increased but was far less than the sum of the individual registration errors. Registration errors due to dehydration of internal prostate tissue components during fixation are unavoidable and will cause shrinkage of the prostate specimen compared to the prostate in the human body [22]. Schned et al. estimated that the in vivo tumor volume was on average 12.4% higher than that measured from prostatectomy pathology [40]. Our own measurements obtained from well-delineated BPH nodules agree with this assessment. The less-than-perfect overlap coefficients (Dice and Jaccard < 1) obtained from BPH nodules can therefore in part be attributed to fixation-induced BPH shrinkage. Conversely, tumor shrinkage may have been smaller than in normal glandular tissue due to higher cellularity of tumors. Nonetheless, tumor shrinkage bears the potential to further underestimate tumor volumes on imaging relative to pathology.

As recently reviewed, the data on tumor volume estimates by MRI relative to histology are mixed; however, most published data report an underestimation of tumor volumes by mpMRI [32]. What may appear surprising in our study is the extent of the underestimation compared to prior reports. When considering the technical improvements of mpMRI over the last decade with the availability of 3 T MRI, one would expect improving accuracy in tumor border delineation. However, considering three prior reports comparing mpMRI at 3 T with whole-mount histology [27, 41, 42], two of which used 3D molds for sectioning, underestimation of tumor volumes between 7 and 80% were noted. The lowest underestimation was found by Turkbey et al. [27]; however, their study was limited due to tumor volume measurements based on an ellipsoid formula rather than a true volumetric measurement, which based on our data significantly overestimates lesion volumes. Also, gross sectioning of prostate specimen was performed manually in almost half of the subjects, which is technically extremely difficult to be done at 4-mm intervals thereby introducing errors [22]. Furthermore, they did not account for misplacement of tumor borders on histology relative to in vivo MRI as done in our study. Le Nobin et al. did not use 3D molds to facilitate registration of MRI and histology, nonetheless reporting on average 57% greater tumor volumes on histology vs. mpMRI evaluating retrospectively 46 lesions of overall smaller volume (mean 1.1 mL) compared to our cohort [42]. More recently, Priester et al. studied 114 subjects undergoing prostatectomy using individually 3D-printed molds for sectioning. The mean volume (2.5 mL) of 107 clinically significant prostate cancers was similar to our data, and the majority of tumors were graded as Gleason ≥ 3+4 [41]. Although they did not determine volumetric similarity coefficients, their overall results are comparable to our study. Eighty percent of the cancer volume from matched tumors was outside of the MR-defined tumor volume, a result that is very similar compared to our own data. The mean HD between histology and mpMRI of clinically significant tumors was 15.6 mm, which was even worse compared to our data (HD of 9.3–10.7 mm for human readers).

A potential reason for the observed disparities of tumor volume estimates obtained from in vivo imaging vs. histology from prostatectomy specimen could have been related to interim tumor growth, particularly as 5 of 10 patients underwent surgery roughly 5–7 months after imaging. While absolute volume growth rates of untreated Gleason ≥ 3+4 cancers are unknown, adenocarcinomas of the prostate are typically progressing slowly [43]. Supporting evidence for slow progression of intermediate risk prostate cancer comes from the observation arms of the Scandinavian Prostate Cancer Group Study Number 4 [44] and the Protect trial [45]. Indeed, 9 of 10 patients included in our cohort harbored intermediate-risk (Gleason 3+4 or 4 + 3) cancer. In one patient with high-risk cancer (Gleason 4+5) and the longest time interval (204 days) between imaging and prostatectomy (patient 9 in Table 1 and Fig. 4), the underestimated tumor volume was however well within the range of other segmentations. Therefore, we consider interval tumor growth not to be a significant contributing factor for the observed underestimation of the tumor volumes.

While the underestimation of tumor volumes on PET can be attributed to technical aspects of PET imaging including partial volume effects limiting image resolution [46], the reasons for underestimated tumor volumes on MRI are less clear. One possibility could be an inhomogeneity of Gleason pattern within a given tumor, where the “detectable” center of such lesions has higher cellular density compared to the periphery of lesions. Since Gleason 3+3 cancers are often missed on MRI [41] and lack focal elevated ¹¹C-choline [46] and ¹⁸F-choline uptake on PET above background [21], we investigated whether mpMRI may only have identified certain pockets of more aggressive disease (Gleason ≥ 3+4), while missing areas of low-grade disease within inhomogeneous prostate cancers. However, a specific analysis of all Gleason ≥ 3+4 cancers did not identify any inhomogeneity of Gleason pattern within individual lesions that could explain the profound underestimation of tumor volumes. Nonetheless, missed tumors are often small and have lower grade than visible tumors on MRI. In addition, non-visible disease may display benign prostatic glandular tissue intermixed with carcinoma [47]. Not only the cellular density or its distribution, but also the cellular composition, particularly the amount of interstitial stromal space, has an effect on DWI and derived ADC maps [48]. Since DWI plays an important role for visual identification of prostate cancer on mpMRI, this phenomenon could explain—at least in part—the underestimation of tumor volumes and underlying uncertainties about tumor borders.

Visual tumor segmentation based by mpMRI and semi-automatic segmentation based on ¹⁸F-choline uptake measures significantly underestimated the true volume of Gleason ≥ 3+4 prostate cancer, and substantial safety margins (up to 15 mm in our series) are required to include all diseases. Combining MR-based human segmentations with a semi-automated thresholding approach based on ¹⁸F-choline PET reduced the necessary safety margin to a maximum of 9 mm. While further work in a larger cohort is needed to validate these findings, this approach might contribute to improving the outcome of focal therapies of significant prostate cancer.