Deep learning-based lesion characterization and outcome prediction of prostate cancer on [ 18 F]DCFPyL PSMA imaging

doi:10.21203/rs.3.rs-5243056/v1

Download PDF

Research Article

Deep learning-based lesion characterization and outcome prediction of prostate cancer on [ 18 F]DCFPyL PSMA imaging

https://doi.org/10.21203/rs.3.rs-5243056/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

This study aimed to develop deep learning (DL) models for lesion characterization and outcome prediction in prostate cancer (PCa) patients using Prostate-Specific Membrane Antigen (PSMA) PET/CT imaging.

Methods

The study included 358 confirmed PCa patients who underwent [¹⁸F]DCFPyL PET/CT imaging. Patients were divided into training and internal test sets (n = 275), prospective test set (n = 64), and external test set (n = 19). Lesions were evaluated using PSMA-Reporting and Data System (RADS) scores, malignancy classification, treatment response and survival prediction, followed by DL models trained for each of these tasks. The performance of multi-modality (PET + CT) models was compared to single-modality models, with the best models from the internal and prospective test sets applied to the external test set.

Results

The input concatenation model, incorporating both PET and CT data, demonstrated the highest performance across all tasks. For PSMA-RADS scoring, the area under the receiver operating characteristic curve (AUROC) was 0.81 (95% CI: 0.80–0.81) for the internal test set, 0.72 (95% CI: 0.69–0.75) for the prospective test set, and 0.68 (95% CI: 0.68–0.69) for the external test set. For malignancy classification, the model achieved AUROCs of 0.79 (95% CI: 0.78–0.80), 0.70 (95% CI: 0.68–0.71), and 0.62 (95% CI: 0.61–0.63) in the internal, prospective, and external test sets, respectively. The AUROC for treatment response prediction was 0.74 (95% CI: 0.73–0.77) for the internal test set, 0.70 (95% CI: 0.67–0.72) for the prospective test set, and 0.72 (95% CI: 0.70–0.73) for the external dataset. The C-index for survival was 0.58 (95% CI: 0.57–0.59), 0.60 (95% CI: 0.60–0.63) and 0.59 (95% CI: 0.57–0.62) in the internal, prospective, and external test sets, respectively.

Conclusions

The DL model utilizing input concatenation of PET and CT data outperformed single-modality models in PSMA-RADS scoring, malignancy classification, treatment response assessment, and survival prediction, highlighting its potential as a clinical tool.

PSMA PET/CT

deep learning

prostate cancer

PSMA-RADS

prognosis

Prostate cancer (PCa) is a prevalent and complex malignancy affecting men globally, ranking as the second leading cause of cancer-related death among adult males in the United States. It accounts for 29% of newly diagnosed male cancers, with a notably low survival rate in the metastatic stage [1]. Effective management of PCa hinges on timely and accurate lesion detection, characterization, and monitoring, along with assessing treatment response and survival prognosis. Prostate-Specific Membrane Antigen (PSMA) has emerged as a critical biomarker for PCa, particularly in its advanced stages [2]. Elevated PSMA expression correlates with worse overall survival and a higher risk of recurrence following curative surgery [3, 4]. PMSA-targetting positron emission tomography (PET) agents hold significant promise in identifying metastatic lesions [5–7]. Additionally, PSMA PET/CT has shown efficacy in detecting biochemically recurrentce of PCa, even at low serum prostate-specific antigen (PSA) levels [7, 8], and aids in identifying sites of recurrence and delineating lesions for targeted biopsy [6].

The PSMA Reporting and Data System (PSMA-RADS), a 5-point-scale framework, is utilized by clinicians to categorize findings from PSMA-PET scans, quantifying the likelihood of PCa lesions [9]. However, this process traditionally relies on the visual interpretation of images by experienced radiologists, making it time-consuming, expertise-dependent, and prone to high interpretive variance. Indeterminate lesions (PSMA-RADS-3A and PSMA-RADS-3B) present additional challenges in distinguishing between benign or malignant cases. Despite advances in PCa treatment [10] and growing survival data from retrospective analyses and prospective studies [11–13], predicting treatment response and patient outcome at the lesion level remains difficult.

Deep learning (DL), a branch of artificial intelligence, is transforming medical image analysis by automating time-consuming tasks, enhancing diagnostic accuracy, and mitigating interobserver variability [14]. Convolutional neural network, a type of DL architecture designed for image analysis, are now integral to medical image processing [15, 16]. Previous studies, including a DL and radiomics framework we developed for PCa lesion classification using PSMA-RADS scores [17], and a 3D CNN by Kendrick et al. for survival prediction in biochemically recurrent PCa [7], demonstrate eh potential of DL. However, these studies have limitations:. lesions lacked imaging or pathological confirmation by follow-up, models were based solely on PET data, and no prospective or externa validation was conducted. Additionally, there is limited evidence supporting image-derived treatment response and outcome prediction.

To address these graps, this study aimed to develop comprehensive DL models that classify lesions based on PSMA-RADS scores with imaging/pathology follow-up, while also predicting treatment responses and patient survival in PCa using PSMA PET/CT imaging and clnical data.

Study Design

This retrospective study was approved by the institutional review board at Johns Hopkins Hospital with a waiver of written informed consent (IRB00349673). Data were collected from three groups of PCa patients who underwent [¹⁸F]DCFPyL PET/CT imaging: 275 patients in Cohort 1 (January 2015 to December 2018) enrolled in a research setting from a previously described study [17]; the first consecutive 64 patients in Cohort 2 (October 2021 to November 2022) following clinical approval of the radiotracer at our institution; and 19 patients from an external institution in Cohort 3 (January 2017 to December 2023). All patients had [¹⁸F]DCFPyL PET/CT imaging and pathological/clinical confirmation of PCa diagnosis. Cases with poor image quality, artifacts, or absence of uptake lesions were excluded, resulting in a final dataset for PSMA-RADS scoring of 238 patients from Cohort 1, 36 patients from Cohort 2, and 19 patients from Cohort 3. In Cohort 1, patients were randomly assigned to training (n = 172) and internal test sets (n = 66), while Cohort 2 and Cohort 3 served as prospective and external test sets, respectively.

Each patient’s chart was reviewed for pathology reports and/or follow-up imaging to categorize lesions on the initial PSMA PET/CT as benign or malignant and to assess treatment response. Of 191 patients with available follow-up data for malignancy classification, 9 were excluded due to lesion removal, preventing assessment of lesion progression. This left 182 patients eligible for treatment assessment evaluation and survival analysis. A detailed flowchart for patient inclusion and exclusion is shown in FIGURE 1. Additionally, clinical variables such as age, race, height, weight, body mass index, PSA levels, Gleason scores, imaging indications, relapse, survival status, therapeutic lines and interval between baseline and follow-up scans were collected. These clinical variables were compared across datasets (SUPPLEMENTARY TABLES 1–4).

Lesion segmentation

All lesions were segmented in the axial plane using Mirada DBx software on a per-slice basis, as previously published [17]. Two radiologists (LZ, YM) performed manual lesion segmentation, which was subsequently reviewed and revised as needed by a third radiologist (HB) as needed.

PSMA-RADS scoring and malignancy evaluation

Each lesion was assigned a PSMA-RADS score (PSMA-RADs version 1.0) [9] by two radiologists (HW, LZ). Disagreements were resolved by a third radiologist (YM). Lesions were grouped based on their PSMA-RADS score, with PSMA-RADS-1 and − 2 lesions classified into one group and PSMA-RADS-3,-4, and − 5 lesions into another for binary PSMA-RADS classification. The training, internal, prospective and external test sets comprised 2125, 915, 300, and 223 lesions, respectively.

For lmalignancy evaluation, two radiologists (LZ and HB) labeled lesions as malignant based on pathology confirmation or follow-up imaging (MRI/CT/PET) showing size changes greater than 2 mm, newly enlarged lymph nodes (over 10 mm), or bone destruction/formation [18]. Lesions not meeting these criteria were labeled as benign. For the benign versus malignant classification task, the training, internal, prospective and external test sets comprised 1217, 370, 168 and 210 lesions, respectively.The distribution of PSMA-RADS scores and malignancy categories across the datasets is detailed in SUPPLEMENTARY TABLE 5.

Lesion treatment response and rurvival evaluation

For treatment response assessment, two radiologists (LZ and HB) labeled lesions as progressive if follow-up imaging (CT, MRI, or PET/CT) showed enlargement of over 2 mm, newly enlarged lymph nodes (over 10 mm) or bone destruction/formation [18]. Lesions that remained stable or shrunk by more than 2 mm were labeled as non-progressive. Surival status at the endpoint (the date of follow-up) for each patiens was also collected.

Deep learning model training and visualization

The models were implemented using Pytorch [19] and MONAI [20], and trained on a NVIDIA GeForce RTX 3090 GPU. For PSMA-RADS score classification and benign-malignant categorization, seven models were developed using PET and CT data (SUPPLEMENTARY FIGURE 1). Two models used single modality (PET or CT) input, while five models applied fusion strategies to combine the modalities. For the treatment response and survival prediction tasks, three single models were developed using PET, CT data, and clinical data individually. Two additional models combined either PET or CT with clinical data, and one model integrated both modalities with clinical data.

For single modality models, a 3D DenseNet architecture [21] was initialized with one input channel for image features. The model began with a 3D convolutional layer, followed by batch normalization, ReLU activation and max-pooling layers. After feature extraction in this initial layer, six densely connected layers formed the first dense block. Each dense layer consisted of two convolutional layers with batch normalization and ReLU activation in between. This densely connected structure facilitated the propagation and reuse of features across the network, enhancing representational power. Transition blocks, incorporating average pooling, separated the dense blocks. Subsequent dense blocks had 12, 24, and 16 layers, respectively, with transition blocks in between. Following the fourth dense block, the model applied batch normalization, ReLU activation, global average pooling, flattening, and finally, the features entered a linear layer for classification with two output features (Supplementary Fig. 1).

Using this DenseNet framework, three late fusion models were developed. PET and CT data were processed separately using the described architecture, and their features are fused prior to the classification layer using one of three strategies: (1) Multi-Layer Perceptron and Self-Attention [22], (2) Squeeze and Excitation (SE) with Sigmoid Activation [23], or (3) convolution blocks. These strategies were termed Output Transformer, Output SE, and Output Convolution. Two early-fusion models combined the two modalities prior to the first dense block using either (1) 3D convolution or (2) concatenation, termed Input Convolution and Input Concatenation.

The models were trained with AdamW optimizer [24] using a learning rate of 1e-5, a batch size of 10, and 1000 epochs. PET/CT images from both datasets were preprocessed uniformly, including conversion of PET intensities to Standardized Uptake Value corrected for body weight (SUVbw) and normalization to the [0,1] range. Volumes were resampled to a 2 mm slice thickness and cropped to a size of 96×96×96 voxels, focusing on normal tissues, PSMA lesions, and surrounding areas. Final models were selected based on classification accuracy on a training subset.

For PSMA-RADS score classification and benign-malignant categorization, prediction probability scores and Uniform Manifold Approximation and Projection (UMAP)[25] feature reduction analysis were applied. The final layer of the DL model provided prediction probabilities for PSMA-RADS group or malignancy status, and the argmax function determined the predicted class. Probability scores for PSMA-RADS-3,-4,-5 and malignancy class were visualized in UMAP space.

The treatment response model incorporated both image-based and clinical data. For image-based inputs, PSMA segmentation masks were used to create masked CT or PET images, which were processed through a DenseNet architecture pre-trained on ImageNet, with four additional predictive layers. Clinical data were processed through a neural network with dense layers of 16, 32, and 2 nodes, utilizing 13 clinical variables to differentiate between non-progressive and progressive treatment outcomes. The final model combined predictions from image-based models (CT, PET or both) and the clinical data-based models through a weighted sum.

For survival prediction, a time-to-event model was employed to estimate the probability of reaching critical outcomes (i.e., death). The image-based model extracted 256-dimensional features from the dense layer of the treatment prediction model, while the clinical data model used the same 13 clinical variables as before. These features were fed in a survival forest model to calculate survival probability scores. The final survival probability for each patient was derived from the weighted sum of the image-based and clinical-data-based risk scores, and the model’s performance was evaluated using time-to-event analysis across different configurations of image and clinical data models.

Statistical analysis

Statistical analysis and data preprocessing were performed using Python v. 3.10.12. For PSMA-RADS score classification, several performance metrics were calculated including accuracy, area under the receiver operating characteristic curve (AUROC), weighted F1 score, precision, recall, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). The best-performing model in the internal and prospective test sets, based on accuracy and AUROC, was selected for evaluation on the external test set. Bootstrap resampling (1000 samples) was employed to calculate 95% confidence and tolerance intervals for ROC curves and accuracy.

The impact of demographic and clinical variables on classification accuracy was assessed using Chi-square tests for categorical variables and t-tests or ANOVA for continuous variables. Consistency between PSMA-RADS scores, model outputs, and ground truth malignancy was evaluated using Intra-class Correlation Coefficient (ICC).

For survival prediction models, accuracy was evaluated with the concordance index (C-index) to account for right-censored data, correlating treatment response with predicted survival probabilities. Patient stratification for survival analysis was performed using the Kaplan-Meier method, with statistical significance assessed via the log-rank test, based on predicted survival probabilities. Time-dependent ROC-AUC was calculated to evaluate survival prediction models over time, and precision-recall curves were generated to quantify precision, recall, and F1 scores. Statistical significance was defined as P < 0.05.

Patient characteristics and dataset summary

There was no significant differences in patient height, body mass index, PSA levels, Gleason scores, survival status, therapeutic lines across the training, internal, prospective and external test sets for PSMA-RADS scoring, malignancy classification, treatment response and survival analysis (all P > 0.05). However, the prospective test set had significantly more patients whose indications were for primary staging and therapeutic response assessment compared to other test sets (P < 0.001). The three cohorts in the PSMA-RADS classification task showed variations in age (P < 0.001), race (P = 0.003), indications (P < 0.001), relapsed status (P < 0.001) and treatment types (P < 0.001). Similarly, the benign versus malignant classification task exhibited differences in age (P = 0.007), weight and race (P < 0.001), indications (P < 0.001), relapsed status (P < 0.001) and treatment types (P < 0.001). For treatment response and survival analysis, significant differences were noted in age (P = 0.011), race and weight (P < 0.001), indications (P < 0.001), relapsed status (P < 0.001), and interval between baseline and follow-up scans (P < 0.001). Detailed patient characteristics and dataset summary are provided in the Supplemenatry Materials (SUPPLEMENTARY TABLES 1–4).

Across all datasets, lesions scored as PSMA-RADS-5 had the highest percentage of confirmed malignancy (58%), followed by PSMA-RADS-4 (28%) and PSMA-RADS-3 (8%) (FIGURE 2A). Lesions located in the prostate (96%) and bone (49%) were more likely to be malignant (FIGURE 2B). The distribution of lesions across datasets is detailed in SUPPLEMENTARY TABLE 5.

Model performance for PSMA-RADS scoring

For PSMA-RADS classification, the input concatenation model outperformed all others, achieving the highest accuracy of 0.73 (95% CI: 0.73–0.74), AUROC of 0.81 (95% CI: 0.80–0.81) (FIGURE 3A), and weighted F1 score of 0.73 (95% CI: 0.73–0.74) in the internal test set. In the prospective test set, the input concatenation model maintained the highest performance with an accuracy of 0.77 (95% CI: 0.76–0.79), AUROC of 0.72 (95% CI: 0.69–0.75) (FIGURE 3B), and F1 score of 0.79 (95% CI: 0.78–0.80). In the external test set, the model achieved an accuracy of 0.74 (95% CI: 0.72–0.76), AUROC of 0.66 (95% CI: 0.66–0.67) (FIGURE 3C), and weighted F1 score of 0.72 (95% CI: 0.71–0.74).

Model performance for malignancy classification

For benign versus malignancy classification, the input concatenation model achieved the highest performance in the internal test set, with an accuracy of 0.78 (95% CI: 0.77–0.79), AUROC of 0.79 (95% CI: 0.78–0.80) (FIGURE 3D), and weighted F1-score of 0.78 (95% CI: 0.77–0.79). In the prospective test set, the model reached an accuracy of 0.70 (95% CI: 0.68–0.71), AUROC of 0.76 (95% CI: 0.74–0.77) (FIGURE 3E), and weighted F1-score of 0.69 (95% CI: 0.67–0.71). In the external test set, the model achieved an accuracy of 0.62 (95% CI: 0.61–0.63), AUROC of 0.68 (95% CI: 0.68–0.69) (FIGURE 3F), and weighted F1-score of 0.62 (95% CI: 0.61–0.63). For lesions scored as PSMA-RADS-3 across all test sets, the input concatenation model correctly predicted 70% of benign lesions and 78% of malignant lesions, using pathology and/or imaging follow-up as the gold standard (FIGURE 2C). Additional details on lesion distribution, model performance, prediction probability scores, and UMAP feature space can be found in the Supplementary Materials. Detailed model performance for PSMA-RADS score and malignancy classification, as well as the ICC values between ground truth and PSMA-RADS scores, are presented in SUPPLEMENTARY TABLES 6–9. UMAP analysis of PSMA-RADS score and malignancy classification is shown in SUPPLEMENTARY FIGURE 2.

Model performance for treatment response and survival prediction

Combining PET and CT data for treatment response prediction yielded the best performance, with an AUROC of 0.74 (95% CI: 0.73–0.77) (FIGURE 4A) and an F1-score of 0.78 (95% CI: 0.77–0.79) in the internal test set. In the prospective test set, the model achieved an AUROC of 0.70 (95% CI: 0.67–0.72) (FIGURE 4B) and an F1-score of 0.70 (95% CI: 0.65–0.69). In the external test set, the model reached an AUROC of 0.72 (95% CI: 0.70–0.73) (FIGURE 4C) and an F1-score of 0.71 (95% CI: 0.70–0.73). The integrated PET/CT model significantly outperformed the single-modality models (P < 0.001).

For survival prediction, the fused PET/CT model also outperformed single-modality and clinical data-based models (P < 0.001), achieving a C-index of 0.58 (95% CI: 0.57–0.59) (FIGURE 4D) and an F1-score of 0.79 (95% CI: 0.78–0.81) in the internal test set. In the prospective test set, the model achieved a C-index of 0.60 (95% CI: 0.60–0.63) (FIGURE 4E) and an F1-score of 0.82 (95% CI: 0.80–0.82). For the external test set, the model achieved a C-index of 0.59 (95% CI: 0.57–0.62) (FIGURE 4F) and an F1-score of 0.81 (95% CI: 0.78–0.83). Detailed results for treatment response and survival prediction are presented in SUPPLEMENTARY TABLES 10–11. Kaplan-Meier curves for risk stratification are shown in SUPPLEMENTARY FIGURE 3.

In this study, we developed DL models that can reliably and accurately classify lesions according to PSMA-RADS score categories on PSMA-PET/CT in patients with PCa, demonstrating their potential as important clinical tools in PCa evaluation and management. Unlike our previous study [17], we confirmed the benign or malignant nature of each lesion through pathology and/or imaging follow-up. Our multi-modality input concatenation model outperformed single-modality models in both the internal and prospective test cohorts, with the exception of PSMA-RADS classification in the external test set, where the PET single model achieved superior performance. This finding underscores the importance of combining functional and structural information for enhanced lesion characterization. Additionally, our multi-modality model successfully predicted lesion-level treatment responses and survival based on baseline PSMA PET/CT imaging and clinical variables, potentially aiding in the early identification of patients at high risk for disease progression.

Standardized frameworks for PET/CT interpretation, such as PSMA-RADS, are valuable not only in diagnosis but also for selecting and monitoring radioligand therapy [26]. In agreement with previous studies [27], lesions with PSMA-RADS scores of 3 or higher, particularly those located in the bone and prostate, were more likely to be malignant, confirming the utilitiy of PSMA-RADS. However, equivocal lesions (e.g., PSMA-RADS-3) continue to present a diagnostic challeng due to the lack of definite indicators of benignity or malignancy. Previous reports indicate that only a minority of equivocal lesions on [¹⁸F]DCFPyL PET/CT are diagnosed as true metastases [18, 28], and studies have noted false-positive uptake in bone with PSMA tracers such as [¹⁸F]PSMA1007 and [⁶⁸Ga]PSMA11, particularly in the ribs [29]. Our study addresses this gap by evaluating the reliability of DL modesl in characterizing lesions based on PSMA-RADS scores, using pathology and/or imaging follow-up as the gold standard. Notably, our model performed well in predicting malignancy in PSMA-RADS-3 lesions, which are often challenging for radiologists [30, 31]. This improved classification of equivocal lesions offers a potential solution to the diagnostic uncertainty of PSMA-RADS-3 and can assist clinicians in making more informed decisions regarding patient management, such as whether further diagnostic procedures or targeted therapeutic interventions are needed. By identifying lesions that warrant closer monitoring, the model can contribute to more personalized treatment strategies and improved patient outcomes.

Previous studies on PCa treatment response and prognosis have predominantly focused on correlating clinical characteristics [32, 33] or radiological features [34] with treatment outcomes. However, these studies often rely on one-dimensional data, limiting their ability to capture individual variations in response to different treatment regimens. A previous study using a 3D U-Net architecture with multi-modal data for survival analysis was retrospectively and did not include treatment response prediction [35]. Similarly, studies on treatment response have larged focused on PSA levels [33, 36] or total tumor volume [37, 38], neglecting changes in individual lesion size or characteristics observed on imaging. Our study addresses these limitatons by integrating PET and CT imaging data with clinical information to develop lesion-level predictive models for treatment response, providing a more granular approach to precision management of individual lesions. This is especially useful for identifying lesions that may benefit from biopsy or targed therapy. However, it is important to acknowledge that treatment response and survival outcomes are influenced by numerous factors [39, 40] beyond those captured by imaging and clinical variables alone. Therefore, further validation and refinement of our models are necessary to optimize their clinical utility.

A limitation of this study is that not all lesions were confirmed by histopathology. The small size of some lesions, especially equivocal ones, made biopsy challenging. As a result, follow-up imaging was often used as the alternative to determine lesion status [41, 42]. Additionally, despite the inclusion of more lesions and an external validation dataset compared to previous work, future studies will benefit from multi-institutional datasets with larger patient cohorts to enhance the robustness and generalizability of our findings. Another limitation is the use of manual segmentation by radiologists as input for the classification network, rather than an end-to-end pipeline that incorporates both segmentation and classification. Furthermore, the heterogeneity of treatment methods in our cohort may have affected the robustness and generalizability of our results. Predicting treatment response and survival for specific treatment regimens may provide more clinically relevant insights. Addressing these limitations will be the focus of future work.

In conclusion, this study successfully developed a DL framework for lesion-level PSMA-RADS score and malignancy classification based on PSMA PET/CT images. Additionally, it is capable of predicting lesion-level treatment response and stratifying patients according to their risk of progression, making it a potential tool for PCa monitoring and informing clinical decision-making.

PCa

Prostate cancer

PSMA

Prostate-Specific Membrane Antigen

PET

positron emission tomography

PSA

prostate-specific antigen

PSMA-RADS

PSMA Reporting and Data System

Deep learning

SUVbw

Standardized Uptake Value corrected for body weight

UMAP

Uniform Manifold Approximation and Projection

AUROC

area under the receiver operating characteristic curve

NPV

negative predictive value

PPV

positive predictive value

ICC

Intra-class Correlation Coefficient

Availability of data and materials

Data from this study are available from the corresponding principal investigator (PI) upon reasonable request.

Acknowledgments

This study was supported by NIH grants U01CA140204, T32EB006351, EB024405, CA134675, and P30CA006973, as well as the Science and Technology Innovation Program of Hunan Province (2021RC5003) and the Ministry of Science and Technology Foreign High-End Experts Introduction Project (G2022161002L).

Author information

Department of Radiology and Radiological Science, Johns Hopkins University School of Medicine, Baltimore, MD, USA

Linmei Zhao, Maliha Imami, Yuli Wang, Wen-Chi Hsu, Ruohua Chen,Andrew F. Voter, Alireza Amindarolzarbi, Lily Kwak,Daniel Kargilis, Shadi Afyouni,Andrei Gafita, Junyu Chen,Xin Li,Jeffrey P. Leal,Yong Du,Harrison X. Bai

Department of Radiology, Xiangya Hospital, Central South University, Changsha, Hunan, China

Yitao Mao,Jingyi Tang,Weihua Liao

Molecular Imaging Branch, National Cancer Institute, NIH, Bethesda, MD, USA

Esther Mena, Peter L. Choyke

School of Informatics, Hunan University of Chinese Medicine, Changsha, China.

Yang Li

Department of Radiology, Second Xiangya Hospital, Central South University, Changsha, Hunan, China

Jing Wu

Department of Diagnostic Imaging, Rhode Island Hospital, Providence, RI, USA

Bi, Lulu, Zhicheng, Jiao

Department of Medical Imaging and Intervention, Chang Gung Memorial Hospital at Linkou, Taoyuan, Taiwan

Wen-Chi Hsu,Gigin Lin

Department of Radiology, University of North Carolina at Chapel Hill, NC, USA

Steven P. Rowe

Department of Radiology, UT Southwestern Medical Center, Dallas, TX, USA

Martin G Pomper

Contributors

L.Z, Y.M, J.T, J.W, A.F.V, A.A acquired data, E.M and P.L.C contributed image and clinical information for external test set. L.Z, W.C.H, S.A, A.G, R.C contributed to interpreting data, M.I, Y.W, Y.L contributed to analyzing and L.Z, M.I, and Y.W contributed to drafting the manuscript, L.Z, H.X. B, S.P.R, M.G.P, L.K, B. L, D.K, A.G, J.P. L critically contributed to revising the manuscript, Z.J, Y.D, G.L, W.L contributed to enhancing its intellectual content; and H.X.B approved the ﬁnal content of the manuscript. All authors read and approved the final manuscript.

Corresponding author

Correspondence to Harrison X. Bai.

Ethics declarations

Ethics approval and consent to participate

This retrospective study was approved by the institutional review board at Johns Hopkins Hospital and written informed consent was waived (IRB: CIR00349673).

Consent for publication

Not applicable.

Competing interests

Martin Pomper is a co-inventor on a U.S. patent covering [¹⁸F]DCFPyL and is entitled to a portion of licensing fees and royalties generated by this technology; this arrangement was approved by Johns Hopkins University in accordance with its conflict-of-interest policies. No other potential conflict of interest relevant to this article was reported.

Siegel RL, et al. Cancer statistics, 2023. CA Cancer J Clin. 2023;73(1):17–48.
Hofman MS, Antigen P-SM. The Target of the Decade, from Biochemical Recurrence to Widespread Adoption. J Nucl Med. 2020;61(Suppl 2):S246–7.
Ghosh A, Heston WD. Tumor target prostate specific membrane antigen (PSMA) and its regulation in prostate cancer. J Cell Biochem. 2004;91(3):528–39.
Wright GL Jr., et al. Expression of prostate-specific membrane antigen in normal, benign, and malignant prostate tissues. Urol Oncol. 1995;1(1):18–28.
Afshar-Oromieh A, et al. [68Ga]Gallium-labelled PSMA ligand as superior PET tracer for the diagnosis of prostate cancer: comparison with 18F-FECH. Eur J Nucl Med Mol Imaging. 2012;39(6):1085–6.
Maurer T, et al. Current use of PSMA-PET in prostate cancer management. Nat Rev Urol. 2016;13(4):226–35.
Kendrick J, et al. Fully automatic prognostic biomarker extraction from metastatic prostate lesion segmentations in whole-body [(68)Ga]Ga-PSMA-11 PET/CT images. Eur J Nucl Med Mol Imaging. 2022;50(1):67–79.
Johnsson K, et al. Analytical performance of aPROMISE: automated anatomic contextualization, detection, and quantification of [(18)F]DCFPyL (PSMA) imaging for standardized reporting. Eur J Nucl Med Mol Imaging. 2022;49(3):1041–51.
Rowe SP, et al. PSMA-RADS Version 1.0: A Step Towards Standardizing the Interpretation and Reporting of PSMA-targeted PET Imaging Studies. Eur Urol. 2018;73(4):485–7.
Telli T, et al. Prognostic factors of overall and prostate-specific antigen-progression-free survival in metastatic castration-resistant prostate cancer patients treated with (177) Lu-PSMA-617. A single-center prospective observational study. Prostate. 2023;83(8):792–800.
Sumanasuriya S, De Bono J. Treatment of Advanced Prostate Cancer-A Review of Current Therapies and Future Promise. Cold Spring Harb Perspect Med, 2018. 8(6).
Henriquez I et al. Current and Emerging Therapies for Metastatic Castration-Resistant Prostate Cancer (mCRPC). Biomedicines, 2021. 9(9).
Meyrick D, et al. Real-World Data Analysis of Efficacy and Survival After Lutetium-177 Labelled PSMA Ligand Therapy in Metastatic Castration-Resistant Prostate Cancer. Target Oncol. 2021;16(3):369–80.
Hosny A, et al. Artificial intelligence in radiology. Nat Rev Cancer. 2018;18(8):500–10.
Liu J, Zhao H. Application of convolution neural network in medical image processing. Technol Health Care. 2021;29(2):407–17.
Sarvamangala DR, Kulkarni RV. Convolutional neural networks in medical image understanding: a survey. Evol Intell. 2022;15(1):1–22.
Leung KH, et al. Deep learning and radiomics framework for PSMA-RADS classification of prostate cancer on PSMA PET. EJNMMI Res. 2022;12(1):76.
Yin Y, et al. Follow-up of Lesions with Equivocal Radiotracer Uptake on PSMA-Targeted PET in Patients with Prostate Cancer: Predictive Values of the PSMA-RADS-3A and PSMA-RADS-3B Categories. J Nucl Med. 2019;60(4):511–6.
Ansel J et al. PyTorch 2: Faster Machine Learning Through Dynamic Python Bytecode Transformation and Graph Compilation. in Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2. 2024.
Cardoso MJ et al. Monai: An open-source framework for deep learning in healthcare. arXiv preprint arXiv:2211.02701, 2022.
Iandola F et al. Densenet: Implementing efficient convnet descriptor pyramids. arXiv preprint arXiv:1404.1869, 2014.
Zhang H et al. Self-attention generative adversarial networks. in International conference on machine learning. 2019. PMLR.
Hu J, Shen L, Sun G. Squeeze-and-excitation networks. in Proceedings of the IEEE conference on computer vision and pattern recognition. 2018.
Loshchilov I, Hutter F. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017.
McInnes L, Healy J, Melville J. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018.
Gafita A, et al. Measuring response in metastatic castration-resistant prostate cancer using PSMA PET/CT: comparison of RECIST 1.1, aPCWG3, aPERCIST, PPP, and RECIP 1.0 criteria. Eur J Nucl Med Mol Imaging. 2022;49(12):4271–81.
Kuten J, et al. The significance of equivocal bone findings in staging PSMA imaging in the preoperative setting: validation of the PSMA-RADS version 1.0. EJNMMI Res. 2021;11(1):3.
Bundschuh RA, et al. High Interobserver Agreement on PSMA PET/CT Even in the Absence of Clinical Data. Clin Nucl Med. 2023;48(3):207–12.
Rauscher I, et al. Matched-Pair Comparison of (68)Ga-PSMA-11 PET/CT and (18)F-PSMA-1007 PET/CT: Frequency of Pitfalls and Detection Efficacy in Biochemical Recurrence After Radical Prostatectomy. J Nucl Med. 2020;61(1):51–7.
Vetrone L, et al. Case report: PSMA PET/CT addresses the correct diagnosis in a patient with metastatic prostate cancer despite negative core biopsies and mpMRI. A diagnostic challenge. Front Oncol. 2023;13:1101221.
Gaudiano C, et al. Multiparametric magnetic resonance imaging for the differential diagnosis between granulomatous prostatitis and prostate cancer: a literature review to an intriguing diagnostic challenge. Front Oncol. 2023;13:1178430.
Hartrampf PE, et al. SUV(mean) on baseline [(18)F]PSMA-1007 PET and clinical parameters are associated with survival in prostate cancer patients scheduled for [(177)Lu]Lu-PSMA I&T. Eur J Nucl Med Mol Imaging. 2023;50(11):3465–74.
Moazemi S et al. Estimating the Potential of Radiomics Features and Radiomics Signature from Pretherapeutic PSMA-PET-CT Scans and Clinical Data for Prediction of Overall Survival When Treated with (177)Lu-PSMA. Diagnostics (Basel), 2021. 11(2).
van der Sar ECA et al. Baseline Imaging Derived Predictive Factors of Response Following [(177)Lu]Lu-PSMA-617 Therapy in Salvage Metastatic Castration-Resistant Prostate Cancer: A Lesion- and Patient-Based Analysis. Biomedicines, 2022. 10(7).
Sathekge M, et al. mCRPC Patients Receiving (225)Ac-PSMA-617 Therapy in the Post-Androgen Deprivation Therapy Setting: Response to Treatment and Survival Analysis. J Nucl Med. 2022;63(10):1496–502.
Hofman MS, et al. [(177)Lu]-PSMA-617 radionuclide treatment in patients with metastatic castration-resistant prostate cancer (LuPSMA trial): a single-centre, single-arm, phase 2 study. Lancet Oncol. 2018;19(6):825–33.
de Jong AC, et al. (68)Ga-PSMA PET/CT for Response Evaluation of (223)Ra Treatment in Metastatic Prostate Cancer. J Nucl Med. 2023;64(10):1556–62.
Violet J, et al. Dosimetry of (177)Lu-PSMA-617 in Metastatic Castration-Resistant Prostate Cancer: Correlations Between Pretherapeutic Imaging and Whole-Body Tumor Dosimetry with Treatment Outcomes. J Nucl Med. 2019;60(4):517–23.
Sivakumar S, et al. Comprehensive genomic profiling and treatment patterns across ancestries in advanced prostate cancer: a large-scale retrospective analysis. Lancet Digit Health. 2023;5(6):e380–9.
Kaisary AV. Compliance with hormonal treatment for prostate cancer. Br J Hosp Med. 1996;55(6):359–66.
Thompson J, et al. The role of magnetic resonance imaging in the diagnosis and management of prostate cancer. BJU Int. 2013;112(Suppl 2):6–20.
Expert Panel on Urologic. ACR Appropriateness Criteria(R) Post-Treatment Follow-up of Prostate Cancer: 2022 Update. J Am Coll Radiol. 2023;20(5S):S164–86.

No competing interests reported.

SupplementMaterials.docx

Download PDF

Version 1

posted

You are reading this latest preprint version

Deep learning-based lesion characterization and outcome prediction of prostate cancer on [ 18 F]DCFPyL PSMA imaging

Status:

Version 1

Abstract

Background

Methods

Results

Conclusions

Figures

Background

Methods

Study Design

Lesion segmentation

PSMA-RADS scoring and malignancy evaluation

Lesion treatment response and rurvival evaluation

Deep learning model training and visualization

Statistical analysis

Results

Patient characteristics and dataset summary

Model performance for PSMA-RADS scoring

Model performance for malignancy classification

Model performance for treatment response and survival prediction

Disscussion

Conclusion

Abbreviations

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1