Machine learning predicts conventional imaging metastasis-free survival (MFS) for oligometastatic castration-sensitive prostate cancer (omCSPC) using prostate-specific membrane antigen (PSMA) PET radiomics

Abstract

Purpose:

This study investigated imaging biomarkers derived from PSMA-PET acquired pre- and post-metastasis-directed therapy (MDT) to predict 2-year metastasis-free survival (MFS), which provides valuable early response assessment to improve patient outcomes.

Materials/Methods:

An international cohort of 117 oligometastatic castration-sensitive prostate cancer (omCSPC) patients, comprising 34 from John Hopkins Hospital (JHH) and 83 from Baskent University (BU), were treated with stereotactic ablative radiation therapy (SABR) MDT with both pre- and post-MDT PSMA-PET/CT scans acquired. PET radiomic features were analyzed from a CT-PET fusion defined gross tumor volume ((GTV) or zone 1), and a 5 mm expansion ring area outside the GTV (zone 2). A total of 1748 PET radiomic features were extracted from these zones. The six most significant features selected using the Chi2 method, along with five clinical parameters (age, Gleason score, number of total lesions, untreated lesions, and pre-MDT prostate-specific antigen (PSA)) were extracted as inputs to the models. Various machine learning models, including Random Forest, Decision Tree, Support Vector Machine, and Naïve Bayesian, were employed for 2-year MFS prediction and tested using leave-one-out and cross-institution validation.

Results:

Six radiomic features, including Total Energy, Entropy, and Standard Deviation from pre-PSMA-PET zone 1, Total Energy and Contrast from post-PSMA-PET zone 1, and Entropy from pre-PSMA-PET zone 2, along with five clinical parameters were selected for predicting 2-year MFS. In a leave-one-out test with all the patients, random forest achieved an accuracy of 80 % and an AUC of 0.82 in predicting 2-year MFS. In cross-institution validation, the model correctly predicted 2-year MFS events with an accuracy of 75 % and an AUC of 0.77 for patients from JHH, and an accuracy of 78 % and an AUC of 0.80 for BU patients, respectively.

Conclusion:

Our study demonstrated the promise of using pre- and post-MDT PSMA-PET-based imaging biomarkers for MFS prediction for omCSPC patients.

Introduction

Prostate cancer stands as the most prevalent solid organ cancer in men, with around 288,300 new cases and over 34,700 deaths recorded in the United States in 2023.[1–3]. Androgen deprivation therapy (ADT) is the most widely used to treat metastatic castration-sensitive prostate cancer[4]. However, most patients after ADT will develop castration resistance. With this in consideration, the development of metastasis-directed therapies (MDT) has attracted increasing interest in improving outcomes by treating all active metastases[5–7]. The role of MDT with stereotactic ablative radiation therapy (SABR) in patients with limited metastatic disease (known as oligometastatic) has been evaluated[8–10]. Oligometastasis, proposed by Sam Hellman and Ralph Weichselbaum, is an intermediate state of cancer spread (<=5 metastases) between localized disease and widespread metastases[11,12] that demonstrates a more favorable but variable prognosis. Early work in oligometastatic castration-sensitive prostate cancer (omCSPC) has evaluated treating all active metastases with metastasis-directed therapy (MDT) and has been associated with improved outcomes. MDT has been shown to achieve high rates of local control in treated omCSPC, and some studies have reported improvements in overall survival[13]. Despite this progress, MDT is not a cure for most metastatic prostate cancer patients. A significant current limitation is that outcomes of MDT for omCSPC patients are very heterogeneous across different populations[14,15] with a median progression-free survival (PFS) ranging from 12–41 months[16]. Improving the ability to prognosticate for patients with omCSPC treated with MDT to address this heterogeneity in treatment response is urgently needed to identify patients who need early treatment intensification.

Exact imaging biomarkers for outcome prediction have shown great promise in providing valuable information for treatment optimization to improve outcomes[17–19]. Similarly, developing imaging biomarkers for outcome prediction early on after MDT is critical to identify poor responders so that proper interventions can be pursued to improve their outcomes[20–22]. Imaging biomarkers can be defined using radiomic features, which include quantitative imaging features calculated based on intensity, shape, size, volume, and texture from medical images. In radiomics analysis, the first step requires the identification and segmentation of volumes of the tumor. Subsequently, radiomic features are extracted from the raw image data over the tumor region, which will be used to predict tumor characteristics, such as tumor response or survival. Positron emission tomography (PET) imaging has been used for cancer detection, grading, staging, and monitoring[23–25]. Evidence showed that pre-treatment or post-treatment PET uptake could be used to evaluate treatment response[26,27]. With the advent of molecular imaging, prostate-specific membrane antigen (PSMA)-based PET imaging has demonstrated excellent sensitivity and specificity in detecting occult metastatic diseases[28,29]. However, it is unclear how to interpret the PSMA-PET response of omCSPC following MDT. Recently, PSMA-PET SUVmax changes over time were found to be a useful response biomarker that potentially correlates with 2-year metastasis-free survival (MFS)[14,30,31]. One limitation of this previous study is that only crude SUV metrics were used. Therefore, in this study, we aim to explore advanced radiomic biomarkers from PET images and clinical parameters acquired before and 3–6 months after MDT to build an effective machine-learning model for early response prediction of MDT.

Methods and materials

Patient cohort

We performed an international multi-institutional study involving patients with omCSPC who received stereotactic ablative radiation therapy (SABR) MDT and underwent pre- and post-treatment PSMA-PET/CT. 34 patients from John Hopkins Hospital (JHH) and 83 from Baskent University (BU) were retrospectively acquired in this study with IRB approval. All patients had omCSPC with ≤ 5 metastases on either conventional (CT/radionuclide bone scan) or molecular imaging (PET). All patients received ⁶⁸Ga-PSMA-HBED-CC or ¹⁸F-DCFPyL-PSMA PET/CT before MDT and a follow-up PSMA-PET/CT performed to evaluate treatment response 3–6 months after MDT. Our clinical outcome of interest was conventional imaging defined MFS after the treatment. A 2-year MFS event was defined as the development of a new metastatic lesion detected on conventional imaging excluding pelvic lymph nodes within 2-years following the end of MDT or death of any cause.[32] Metastasis on conventional imaging was chosen as the endpoint as conventionally detected MFS has been established by ICECaP as a surrogate for overall survival. We have also evaluated these data by looking at any radiographic progression (conventional or advanced molecular imaging) with similar results, however, advanced molecular imaging-based progression free survival has not been validated as a surrogate for overall survival and therefore the report of conventional MFS is a more robust clinical endpoint. A flow diagram of the study cohort is shown in Fig. 1. 24 patients were excluded from development and validation of the model due to having less than 2- years of follow-up and/or not experiencing an MFS event. Additionally, one patient with visceral metastases was excluded to maintain consistency.

Fig. 1.

Flow diagram of the study cohort. Patient numbers denoted by N(n).

PET/CT imaging and preprocessing

The image processing workflow is shown in Fig. 2a. PET images from John Hopkins Hospital were performed on Siemens (Biograph128_mCT), with original volumetric dimensions of 400 × 400 × 334, and voxel spacing of 2.03 × 2.03 × 3.30 mm³. CT images were acquired with the original volumetric dimensions of 512 × 512 × 334, voxel spacing of 0.98 × 0.98 × 3.30 mm³, and 120 kVp. PET images from Baskent University were performed on GE medical systems (Discovery STE), with original volumetric dimensions of 192 × 192 × 327, and voxel spacing of 3.65 × 3.65 × 3.27 mm³. CT images were acquired with the original volumetric dimensions of 512 × 512 × 334, voxel spacing of 0.98 × 0.98 × 3.30 mm³, and 140 kVp.

Fig. 2.

(a) The workflow of data processing, radiomic feature extraction, harmonization, combination, selection, and model construction. (b) Depiction of the zones on PET/CT image. The white arrows show that zone 1 is the GTV area, and zone 2 is the 5 mm expansion ring area outside the GTV.

The pre-PSMA-PET/CT images were rigidly and deformable registered to the post-PSMA-PET/CT coordinates by using the VelocityAI software, and PET images were then resampled to the corresponding CT images. Gross tumor volume (GTV) based on PET and CT was identified and manually contoured on each image by two radiation oncologists. A planning target volume (PTV) expansion of 5 mm was performed based on metastasis location, defining zone 1 as the GTV and zone 2 as a 5 mm expansion of zone 1 (ring area outside the GTV). Fig. 2b shows the zones from PET/CT images.

Clinical parameters, features extraction, combination, and selection

Clinical parameters included age, the total number of metastases, untreated metastases, Gleason grade, and pre-MDT prostate-specific antigen (PSA) were assessed based on the recommendations from the physicians. The extraction of radiomic features was conducted using the open-source Python package PyRadiomics.[33] A total of 428 radiomic features were examined, originating from two distinct zones: in Zone 1, we examined 107 features from pre-PSMA-PET $\left({\mathrm{X}}_{pre}^R\in {\mathbb{R}}^{73\times 107}\right)$ and 107 features from $\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{t}-\mathrm{P}\mathrm{S}\mathrm{M}\mathrm{A}-\mathrm{P}\mathrm{E}\mathrm{T}(X_{post}^R\in {\mathbb{R}}^{73\times 107})$. Zone 2 also included 107 features from both pre-PSMA-PET (${\mathrm{X}}_{pre}^R\in {\mathbb{R}}^{73\times 107}$) and post-PSMA-PET (${\mathrm{X}}_{post}^R\in {\mathbb{R}}^{73\times 107}$). These features encompassed a variety of characteristics, including first-order statistics (18), shape (14), gray-level co-occurrence matrix (GLCM) (24), gray-level dependence matrix (GLDM) (14), gray-level run length matrix (GLRLM) (16), gray level size zone matrix (GLSZM) (16), and neighboring gray-tone difference matrix (NFTDM) (5). Additionally, 1320 wavelet features, with 330 features from both pre-PSMA-PET $\left({\mathrm{X}}_{pre}^W\in {\mathbb{R}}^{73\times 330}\right)$ and post-PSMA-PET $\left({\mathrm{X}}_{pre}^W\in {\mathbb{R}}^{73\times 330}\right.)$ in Zone 1, and 330 features from both pre-PSMA-PET (${\mathrm{X}}_{pre}^W\in {\mathbb{R}}^{73\times 330}$) and post-PSMA-PET (${\mathrm{X}}_{pre}^W\in {\mathbb{R}}^{73\times 330}$) in Zone 2 were extracted after wavelet transformation. Given that PSMA PET data was aligned with PSMA CT, we utilized the same settings as those used for analyzing CT data. When running PyRadiomics, we adhered to the IBSI recommendations by implementing feature extraction parameters in the YAML file. These parameters, including imageType (PET), maskType (label), label (1), normalize (true), normalizeScale (1), binWidth (25), sigma (1), voxelArrayShift (1000), resampledPixelSpacing ([2,2,2]), and interpolator(sitkBSpine), were applied to ensure consistent feature extraction. For patients with multiple metastases, we concatenated the radiomics features by selecting the maximum values. Therefore, a total of 428 radiomic features and 1320 wavelet features were extracted from pre-PSMA-PET and post-PSMA-PET, respectively. Each corresponding radiomic feature from pre-PSMA-PET and post-PSMA-PET was combined with the following formula:

$$\mathrm{X}={\mathrm{X}}_{pre}^R\cup {\mathrm{X}}_{post}^R\cup {\mathrm{X}}_{pre}^W\cup {\mathrm{X}}_{post}^W$$

Function ${\chi }^2$ was used to select the most significant six features. For each feature, the value of ${\chi }^2$ will be calculated from the training dataset. The six features with the highest score of ${\chi }^2$ will be selected from the training dataset. The corresponding six features will be selected from the testing dataset.

Model generation and performance evaluation and statistical testing

Several different machine learning models ((Random Forest, Decision Tree, SVM, Naïve Bayesian) were implemented for 2-yr MFS prediction based on radiomics features. The models were tested using both a leave-one-out strategy and cross-validation across the two institutions. For our validation approach, we treat each dataset from a single institution as a retraining set. We then replicate the same processing steps, adhering to the leave-one-out method. To evaluate the performance of the machine learning model, the following quantitative metrics were assessed with the radiographic 2-yr MFS as the ground truth: accuracy, sensitivity, specificity, receiver-operating characteristic (ROC) curve analysis, and the area under the curve (AUC). Furthermore, the model was used to stratify patients into rapid progressors without 2-yr metastasis free survival (MFS) and non-rapid-progressors groups with 2-yr MFS. The Kaplan-Meier (KM) curves of the two groups were plotted using their actual 2-yr MFS and compared to determine their statistically significant differences to evaluate the efficacy of the model stratification.

Result

A total of 92 patients with 238 metastases (153 treated with MDT) were used in the study. Among them, 43 (47 %) patients were confirmed to have experienced distant metastasis within the 2-year timeframe, while 49 (53 %) were confirmed to have survived without metastasis for the same duration. Table 1 shows the summary of patient characteristics. To assess the correlation between clinical parameters and 2-year MFS, statistical analyses were performed by taking p < 0.05 as statistically significant, which was calculated by an independent two-sample t-test.

Table 1.

Summary of metastases characteristics.

Characteristics	Patients from John Hopkins (%)	Patients from Baskent University (%)	P value
Total Patients Number	32 (35 %)	60 (65 %)
Age (y)
Median	67.5	64.5	<0.01
Range	53–84	46–88
Gleason Score
6, 7, 8	28 (88 %)	31 (52 %)	<0.01
9, 10	4 (12 %)	29 (48 %)
Number of lesions
Lesions treated	95 (66 %)	93 (100 %)	<0.01
Lesions untreated	50 (34 %)	0 (0 %)
Metastasis types
Bones	15 (47 %)	46 (77 %)	NA
Node	16 (50 %)	9 (15 %)
Bone and Node	1 (3 %)	2 (3 %)
others	0 (0 %)	3 (5 %)
Pre-MDT PSA
Mean	9.0	13.2	<0.01
Range	0.5–33	0.3–140
Primary treatment
MDT (Surgery or RT)	32 (100 %)	8 (13 %)	NA
ADT	0 (0 %)	12 (20 %)
RT+ADT	0 (0 %)	40 (67 %)
Target volumes
PTV	32 (100 %)	60 (100 %)	NA
Prescription doses
1 × 16, 1 × 18	0 (0 %)	42 (70 %)	NA
2 × 10	0 (0 %)	8 (13 %)
3 × 8, 3 × 9	8 (25 %)	0 (0 %)
5 × 6, 5 × 7	20 (63 %)	7 (12 %)
others	4 (12 %)	3 (5 %)
No 2-yr MFS event at last follow-up	21 (66 %)	28 (47 %)	NA
2-yr MFS event	11 (34 %)	32 (53 %)

We employed a total of 1748 PET radiomic features (428 original and 1320 wavelets) for the prediction of 2-year MFS. During the feature selection process, the function ${\chi }^2$ method demonstrated the most efficacy in identifying the six most significant features: total energy, entropy, and standard deviation from a CT-PET fusion-defined gross tumor volume (GTV) or zone 1 in pre-PSMA-PET/CT scans, total energy and contrast from zone 1 in post-PSMA-PET scans, and entropy from a GTV+5 mm expansion ring area outside or zone 2 in pre-PSMA-PET scans. Among them, entropy, standard deviation, and total energy are the first-order features derived directly from the voxel intensities, while contrast is a Gray Level Co-occurrence Matrix (GLCM) feature that correlates with a greater disparity in intensity values among neighboring voxels. Table 2 shows prediction results using a random forest model. Leave one out testing was first used to evaluate the impact of inputs on the model performance. The key findings are:

1. Pre-and post-imaging: incorporating both pre-PSMA-PET and post-PSMA-PET scans achieved the best prediction accuracy and AUC, reaching 74 % and 0.76, respectively.

2. Tumor microenvironment (TME): combining PET radiomic features from both zone 1 and zone 2 enabled the model to accurately predict 2-year MFS events for 68 (74 %) patients, improved from 66 (72 %) using zone 1 or 60 (65 %) using zone 2 alone.

3. Clinical versus radiomics: the integration of both clinical and radiomic information led to the best model prediction, achieving an accuracy rate of 80 % and an AUC of 0.82.

Table 2.

Detailed prediction results by using different data fusion.

Features Fusion\Random Forest			Accuracy	Sensitivity	Specificity	AUC
Leave-one-out	Impact of Pre- and Post-Imaging	Pre-PSMA-PET (Zone 1 + Zone 2)	66 (72 %)	32 (74 %)	34 (69 %)	0.74
		Post-PSMA-PET (Zone 1 + Zone 2)	60 (65 %)	28 (65 %)	32 (65 %)	0.68
		Pre-PSMA-PET and Post-PSMA-PET (Zone 1 + Zone 2)	68 (74 %)	32 (74 %)	36 (73 %)	0.76
	Impact of Zones	Zone 1 (Pre- + post-)	65 (71 %)	30 (70 %)	35 (71 %)	0.73
		Zone 2 (Pre- + post-)	57 (62 %)	25 (58 %)	32 (65 %)	0.65
		Zone1 and Zone 2 (Pre- + post-)	68 (74 %)	32 (74 %)	36 (73 %)	0.76
	Impact of Clinical parameters	Clinical (Only)	60 (65 %)	28 (65 %)	32 (65 %)	0.67
		Radiomics (Only)	68 (74 %)	31 (72 %)	37 (76 %)	0.76
		Clinical + Radiomics (ALL)	74 (80 %)	34 (79 %)	40 (82 %)	0.82
Cross-validation		Clinical + Radiomics (JHH)	24 (75 %)	6 (55 %)	18 (86 %)	0.77
		Clinical + Radiomics (BU)	47 (78 %)	23 (72 %)	24 (86 %)	0.80

After the model was finalized in the leave-one-out test, we further did cross-institution validation of the model. The function ${\chi }^2$ identified five out of the six features during training with 60 BU patients. As shown in Table 1, random forest predicted 2-year MFS events correctly for 24 (75 % of 32) patients when being trained using 60 BU patients and tested using 32 JHH patients. Vice versa, the function ${\chi }^2$ identified the same six features during training 32 JHH patients, and the model predicted 2-year MFS events correctly for 47 (78 % of 60) patients when being trained using 32 JHH patients and tested using 41 BU patients.

Fig. 3a1 b1 and c1 show the ROC curves of the model prediction of 2-year MFS events with an AUC value above 0.77 in all tests. Furthermore, we used the final random forest model to predict and categorize the patients into two groups: rapid progressors were defined with a 2-year MFS event and non-rapid progressors without a 2-year MFS event. The Kaplan-Meier (K-M) curves of the two groups were plotted based on their actual survival data and compared to evaluate the efficacy of the stratification based on the model. As shown in Fig. 3a2 b2 and c2, the K-M curves for the rapid progressors and non-rapid-progressors groups were well separated with statistical significance in both leave-one-out and cross-validation tests.

Fig. 3.

(a) The ROC curve and K-M curve comparison for all patients stratified by the model in the leave one out test. (b) The ROC curve and K-M curve comparison for JHH patients stratified by the model in the cross-validation test. (c) The ROC curve and K-M curve comparison and for BU patients stratified by the model in the cross-validation test. (Rapid progressors: with a 2-year MFS event. Non-rapid progressor: without a 2-year MFS event.).

Our previous study showed the potential of using SUVmax to predict 2-yr MFS.[30] The Random Forest model demonstrated performance with an accuracy of 64 %, sensitivity of 47 %, specificity of 76 %, and an AUC of 0.66. Here we conducted a comparison between the prediction using radiomics and change in SUVmax. The results are included in the Supplemental Materials. The radiomics model demonstrated an overall prediction accuracy improvement of 10 %, as compared to the SUVmax model.

In addition, we studied the impacts of harmonization through leave-one-out and cross-validation. As expected, our model demonstrated excellent performance through the implementation of feature domain harmonization using ComBat. Detailed leave-one-out and cross-validation results are presented in Supplementary Table S2, highlighting the effect of harmonization. Additionally, we examine the performance of four machine learning models. Supplementary Figures 4a, b, and c provide a comprehensive breakdown of the validation results and the leave-one-out.

Discussion

Our study is novel and significant in the following aspects: (1) We are the first to investigate using radiomic imaging biomarkers from both pre-and post-treatment PSMA-PET/CT together with clinical information for predicting outcomes in omCSPC patients undergoing MDT. (2) Multi-zone feature extraction – most studies have extracted features from a single region, such as the tumor volume defined by GTV, clinical target volume (CTV), or planning target volume (PTV). Recent studies have shown the value of features extracted from the peripheral TME [34]. We extracted features from two distinct zones: zone 1 corresponded to the GTV, and zone 2 encompasses a 5 mm expansion ring area surrounding the GTV. Our results demonstrated the value of the multi-zone analysis around the TME. (3) Multi-institutional validation – our study benefits from patient data collected from two institutions (Johns Hopkins Hospital and Baskent University), which validated the robustness and generalizability of our findings. Ultimately, our cross-institutional validation showcases the potential of our model in the early prediction of 2-year MFS events following MDT treatments for omCSPC patients. This discovery opens up new avenues for closely monitoring and implementing timely treatment interventions for individuals identified with unfavorable prognoses, ultimately leading to improved patient outcomes. Recently, a number of randomized phase 2 studies have shown the promise of MDT to improve outcomes in omCSPC patients. [35–37] Despite these favorable outcomes, a subset of patients progress rapidly and therefore response biomarkers are urgently needed to identify patients at an early time point who require therapeutic intensification. This study demonstrated that 2-year MFS can be successfully predicted (>75 % accuracy) by assessing pre-PSMA-PET and post-PSMA-PET following MDT, indicating that PSMA-PET may contain effective radiomic biomarkers for omCSPC patients. MFS has been shown to be a validated intermediate clinical endpoint that correlates very well with overall survival. Our stratification results also demonstrated the efficacy of this model in stratifying patients into cohorts with different survival curves. Predicting survival curves is also valuable, however, it is currently not practical given the patient size at each institution.

Patients included in this study included those treated on the ORIOLE clinical trial at Johns Hopkins Hospital. ORIOLE was a Phase II randomized trial that compared metastasis-directed therapy vs observation in patients with conventionally detected (CT A/P & NM bone scan) oligometastatic prostate cancer. In this study patients received a pre-treatment PSMA-PET scan however the investigators were blinded to the results of the PSMA-PET. Patients therefore received metastasis-directed therapy (MDT) for all of the conventionally detected metastasis, however, there was a possibility that conventionally occult, PSMA avid lesions would not be treated. In the original report of the ORIOLE trial, patients who had untreated conventionally occult PSMA-avid lesions were associated with worse PFS (HR, 0.26; 95 % CI, 0.09–0.76; P=.006). Given this was a known potential confounder of our results, we included untreated conventionally occult, PSMA avid lesions within our machine learning model. The clinical parameter inputs to the model included the number of untreated lesions. Treatment paradigms vary between institutions in terms of ADT and the prevalence of untreated metastases. However, the model correctly predicted 2-year MFS events with an accuracy of 75 % for patients from JHH, and an accuracy of 76 % for BU patients. Our cross-institution validation findings demonstrate the robustness and efficacy of our model in predicting 2-year MFS, despite these clinical approach variations.

Additionally, our cohorts included patients treated by MDT with or without a very short course of concurrent ADT (median length 2 months) and then followed off systemic therapy until disease progression. The benefit of the post-treatment prediction is that it can guide the design of follow-up treatments. Patients with good prognosis can be kept off ADT after the initial MDT(±ADT) treatments. In contrast, patients with poor prognosis should likely be started/restarted (in the case of those treated with concurrent ADT) on ADT or intensified hormones, i. e., Abiraterone, enzalutamide, etc, shortly after the initial treatments.

The study has several limitations. First, there is a variation in the timing for post-treatment imaging in the patient cohorts. The imaging features may change over time, leading to variability in radiomic features. Standardizing and optimizing the imaging timing can potentially further enhance the prediction accuracy. Second, our study is limited by the sample size of 92 patients after excluding 24 patients without adequate follow up. In future work, we will continue to accrue more patient data to further train and validate our model performance in a larger patient cohort. Third, our study excluded ADT from our model. ADT has the potential to predict the 2-year MFS. However, ADT was not administered to the Johns Hopkins patients due to the ORIOLE trial’s restriction to MDT-only treatments, while it was used for most patients at Baskent University. Fourth, several studies have highlighted the variation between F-based and Ga-based PSMA radiotracers.[38–40] The overall American Joint Committee on Cancer prognostic stage was comparable, with a 92 % similarity. [41] In this study, two types of radiotracers were used due to logistical and clinical considerations: patients at Johns Hopkins were scanned with F-based PSMA, while those at Baskent University were scanned with Ga-based PSMA. However, despite this difference between the tracers, our radiomic-clinical model still achieved promising results in the cross-institution validation, demonstrating its robustness against such variations. Unifying the radiotracers could further improve the prediction accuracy in the future.

Conclusion

Our study demonstrated the promise of using pre- and post-MDT PSMA-PET-based imaging radiomic biomarkers along with clinical parameters for 2-year MFS prediction for omCSPC patients. Imaging biomarkers predictive of 2-yr MFS were identified in both GTV and the ring area outside GTV. Over 75 % prediction accuracy was achieved in the cross-institution validation test. The model provides a valuable tool for prognosis prediction early following MDT, which opens up a unique opportunity for monitoring or treatment interventions for patients identified with poor prognoses to improve outcomes.