Convoluted Neural Network for Detection of Clinically Significant Prostate Cancer on ⁶⁸ Ga PSMA PET/CT Delayed Imaging by Analyzing Radiomic Features

Abstract

Purpose

To assess the utility of convoluted neural network (CNN) in differentiating clinically significant and insignificant prostate cancer in patients with ⁶⁸ Ga PSMA PET/CT-targeted prostate biopsy-proven prostate cancer.

Methods

In this retrospective study, 142 patients with clinical suspicion of prostate cancer were evaluated who underwent ⁶⁸ Ga-PSMA PET/CT imaging followed by ⁶⁸ Ga-PSMA PET/CT-targeted prostate biopsy from the PSMA-avid prostate lesion. Twenty patients with no PSMA-avid lesions were excluded. Local Image Features Extraction (LifeX) software was used to extract radiomic features (RF) from delayed ⁶⁸ Ga-PSMA PET/CT images of 122 patients. LifeX failed to extract radiomic features in 24 patients, and the remaining 98 were evaluated. RFs were fed to an in-built CNN of the software for computation and results were achieved. Patients with Gleason Score ≥ 7 on histopathology were labeled clinically significant prostate cancer (csPCa). The diagnostic values of radiomic features were evaluated.

Results

The csPCa was revealed in 69/98 (70.4%) patients, and insignificant PCa was noticed in 29/98 (29.6%) patients. The software extracted 124 RF from the delayed ⁶⁸ Ga-PSMA PET/CT images. The accuracy of the CNN was 80.7% to differentiate clinically significant and clinically insignificant prostate cancer, with an error percentage (E %) of 19.3%. The sensitivity, specificity, positive predictive, and negative predictive values were 90.3%, 57.7%, 83.6%, and 71.4%, respectively, to detect csPCa.

Conclusion

CNN is a feasible pre-biopsy screening tool for identifying clinically significant prostate cancer and can be used as an adjunct in the initial diagnosis and early treatment planning.

Supplementary Information

The online version contains supplementary material available at 10.1007/s13139-023-00832-3.

Introduction

Prostate cancer is one of the leading causes of cancer mortality worldwide, ranking sixth in the list [1]. Even though prostate cancer incidence is less in India than in high-income nations, the sub-optimal cancer data maintenance and absence of prostate-specific antigen (PSA) screening in many parts of the country mask its actual prevalence [2]. Hence, a robust analysis of the available cancer data and the incorporation of upcoming technology can provide insight into the behavior of the pathology in the Indian sub-continent. The differentiation of clinically significant prostate cancer (csPCa) and insignificant prostate cancer is the cornerstone for managing prostate cancer [3]. ⁶⁸ Ga PSMA PET/CT is shown to be superior in the initial staging of primary prostate cancer to detect metastasis and in the detection of index lesions in the case of multifocal prostatic lesions [4, 5]. There is an upcoming trend in utilizing ⁶⁸ Ga PSMA PET/CT-guided biopsy to establish a prostate cancer diagnosis. PSMA PET-guided prostate biopsy is shown to detect PCa and csPCa in 84% and 80% of patients, respectively, and has proven to be superior to trans-rectal ultrasound-guided prostate biopsy [6]. Developing a screening tool to guide biopsy planning using PSMA PET/CT for nuclear medicine interventionists to standardize the practice is imperative. Artificial intelligence (AI) is a cheap yet complex technology that can provide insights into the available data through enhanced analysis mechanisms. Using RF to differentiate PCa and csPCa has been explored extensively by MRI [7, 8]. Studies on PSMA PET radiomic features analysis are a handful, but their results favor cancer characterization [9–11]. In the present study, we aimed to assess the utility of convoluted neural network (CNN) in differentiating significant and non-significant prostate cancer in patients with ⁶⁸ Ga PSMA PET/CT-guided prostate biopsy-proven prostate cancer.

Materials and Methods

Patient Population

In this retrospective study, we analyzed the PSMA PET/CT data of 142 patients who underwent ⁶⁸ Ga PSMA PET/CT and were followed by PSMA PET/CT-guided prostate biopsy from January 2018 to August 2022.

⁶⁸ Ga PSMA PET/CT Scan Acquisition and ⁶⁸ Ga PSMA PET/CT-Guided Biopsy

All the patients underwent whole-body ⁶⁸ Ga PSMA PET/CT imaging with clinical suspicion of prostate cancer on a dedicated PET/CT scanner (Discovery MIDR, GE Healthcare, USA). In all patients with PSMA expressing prostate lesions with no previous histological diagnosis, PSMA PET/CT-guided biopsy was planned from the tracer-avid prostatic lesion. A limited FOV delayed PET/CT of the pelvic region after urinary bladder voiding was acquired after 2 h of the radionuclide injection before the biopsy for biopsy planning.

Imaging parameters for CT were helical, 120 kVp, 40 mA, 1.375:1 pitch, and 3.75-mm collimation with reconstruction as 1.25-mm-thick sections by using a 512 × 512 matrix. PET emission data was acquired for 3 min. PET images were reconstructed in a 128 × 128 matrix using CT data for attenuation correction and an iterative-ordered subset expectation–maximization algorithm.

The rationale for acquiring a delayed scan was to (i) ensure adequate tracer washout from the bladder for better delineation of the prostate lesion, (ii) reduce the radiation exposure to the interventionist, and (iii) increase the lesion-to-background ratio of tracer uptake [12]. A ⁶⁸ Ga PSMA PET/CT-targeted prostate biopsy was performed by the interventionist. The histopathological reports of the prostate biopsy were collected and recorded. The images of all 142 patients and their recorded data were used for analysis. Of these, 20 patients had no focal lesion in the prostate on PSMA PET/CT imaging and were excluded. The flow chart of the methodology is mentioned in Fig. 1.

Fig. 1.

Flow chart demonstrates the workflow of the study

Radiomic Features Extraction and CNN Training

We analyzed the delayed ⁶⁸ Ga PSMA PET/CT images of patients (n = 122) who underwent PSMA PET/CT-guided prostatic biopsy from PSMA-avid lesion. A Local Image Features Extraction (LifeX v7.2.0, https://www.lifexsoft.org/) freeware was used to extract radiomic features [13]. The images were loaded in LifeX software, and the region of interest (ROI) was drawn over the PSMA-avid prostate lesion by a nuclear medicine physician with more than 10 years of experience (Fig. 2) using a semantic segmentation editor. Automatic radiomic features were extracted using LifeX software, and the values were recorded. No SUVmax intensity threshold was applied in the ROI for feature extraction because (i) ROI was drawn by a trained physician with utmost possible accuracy and (ii) delayed images were used for ROI since they provide better tumor to background delineation. The ROI content was fixed to a minimum of 64 voxels to calculate the higher orders radiomic features. The size of the ROI was adjusted according to the visualized lesion size. The software failed to extract the radiomic features in 24 patients because the size of the ROIs was less than 64 voxels. Hence, finally, 98 patients’ radiomic features were extracted automatically from the drawn 3D ROI by the LifeX software. The ROI was spatially resampled and the intensity discretization was done with fixed 64 bins with a bin size of 0.317. The intensity rescaling was absolute with minimum bound of 0 and maximum bound of 20 as recommended by the software.

Fig. 2.

A Prostate lesion on fused axial PSMA PET/CT image acquired during pre-biopsy planning for ⁶⁸ Ga PSMA PET/CT-guided biopsy. B Demonstration of 3D ROI (pink in color) drawn on the lesion

The patients were grouped into two categories based on their histopathological report: group 1: insignificant prostate cancer with GS < 7 (n1 = 29) and group 2: clinically significant prostate cancer with GS ≥ 7 (n2 = 69). The detailed histology of the 98 patients is summarized in Table 1.

Table 1.

Histological types of prostate cancer

Histology type	Adenocarcinoma NOS	Adenocarcinoma acinar variant
Non-csPCa* (N1 = 29) GS < 7	26	3
csPCa (N2 = 69) GS ≥ 7	63	6

^*csPCa, clinically significant prostate cancer; GS, Gleason Score

An input file was created containing 124 RF of 88 patients and an output file categorizing the patients to either group. The files were then fed to the neural network (two-layer feed-forward network) in MatLab 2020b [14] with a sigmoid hidden layer containing 10 neurons and two softmax output neurons.

The data division was random, training algorithm was scaled conjugate gradient, and performance was assessed using cross-entropy.

CNN was trained and validated by using a training set of 53 (60%) of 88 patients, a validation set of 22 (25%), and a testing set of 13 (15%). After the training, the neural network was tested using an independent test set of 10 patients, of which seven had significant PCa and three had non-significant PCa. The independent dataset was used to test the CNN to eliminate the bias from the training and validation datasets that happens during the training.

Results

All the 98 patients (aged 77.0 ± 6.2 years) who underwent PET/CT-directed biopsy were proven to have prostate cancer. The average PSA level of all the patients was 19.6 ng/ml. A total of 69/98 (70.4%) patients had clinically significant prostate cancer, and these patients had PSA and PSA density of 20.506 ± 15.2 ng/ml and 0.422 ± 0.29 ng/ml². The non-csPCa group (n = 29) had PSA and PSA density of 18.826 ± 14.8 ng/ml and 0.429 ± 0.41 ng/ml², respectively. There was no statistically significant difference between the groups in PSA and PSA densities.

A total of 124 radiomic features using LifeX v7.2.0 software extracted from the regional ⁶⁸ Ga PSMA PET/CT images of the pelvis acquired during the biopsy procedure of these patients are enlisted in supplementary file 1.

CNN for Differentiation of Clinically Insignificant PCa and csPCa

The evaluation of trained CNN was performed. The results of the training and testing of CNN were as follows:

• The training was automated and done by a scaled conjugate gradient with 25 iterations and six validation checks.

• The best validation performance was 0.25569 at epoch 19.

• ROC (receiver operator curve) of all the sets (training, validation, testing, and combined sets) and the final trained CNN’s ROC were plotted (Fig. 3).

• The confusion matrix was obtained. The ability of CNN to differentiate non-significant prostate cancer and clinically significant prostate cancer was 80.7% accurate, and the error percentage (E %) was 19.3% (Fig. 4).

• The CNN differentiated the patients with an accuracy of 70.0% in the independent test set (Figs. 5 and 6)

• The sensitivity, specificity, positive predictive, and negative predictive values were 90.3%, 57.7%, 83.6%, and 71.4%, respectively.

Fig. 3.

ROC training set (a), validation set (b), testing set (c), and combined set (d)

Fig. 4.

Confusion matrix of all the sets. The accuracy of the CNN for the training set, validation set, test set, and the combined set is 88.7%, 63.6%, 76.9%, and 80.7%, respectively

Fig. 5.

Confusion matrix of the trained CNN on independent test dataset. The accuracy is 70.0% of the trained CNN and the error percentage is 30.0%

Fig. 6.

ROC of the trained CNN analyzing the independent dataset

The average SUVmax of the clinically significant prostate cancer group was 27.23 ± 17.58 and of the non-significant prostate cancer group was 18.74 ± 10.30. The SUVmax values of the two groups were not statistically different (p value = 0.54 by independent-samples Mann–Whitney U test). The features which were statistically different (p value < 0.05) between the two groups by independent-samples Mann–Whitney U test were as follows:

• Morphological features: compactness 1, compactness 2, spherical disproportion, sphericity, and asphericity

• Intensity features: energy, intensity histogram range, and maximum histogram gradient

•  Gray scale features: gray-level non-uniformity

• The performance of the CNN remained the same even after the feature selection and testing.

Discussion

Radiomic features analysis carries the capacity to guide oncological clinical decisions, yet this approach’s importance has not gained weightage [15]. A substantial number of studies were conducted on classifying PCa based on the radiomic features of MRI images [7, 8, 16]. Studies have been carried out based on several machine learning approaches. Deep learning methods are more reliable than radiomics-driven feature-based methods for classifying PCa in MRI images [17]. The deep learning application for the classification of PCa in PET/CT images has not yet been explored enough. The present study evaluated the radiomics-driven feature-based CNN classification of PCa in ⁶⁸ Ga PSMA PET/CT images.

ROI is a data structure used in image processing and is prevalent in medical image analysis. Depending on the need, it was used to define an area within the image for further usage [18]. ROI-based radiomics features can be classified using CNN. There are two commonly used ROI types: 2D and 3D. 3D ROIs can provide extra information on cancer characteristics. 3D ROI in MRI images of PCa was used by Mehrtash et al. [19] and classified PCa via a CNN.

Manual ROI-based detection tools are time-consuming, and the segmentation process depends on the expertise of the clinician. Even though the manual ROI-based methods are time-consuming, automated ROI-based methods use segmentation algorithms, which can have poor performance because of inadequacies in the algorithm [8]. In our study, we used a manual 3D ROI-based method. The ROI was defined by a qualified nuclear medicine physician with 15 years of experience in PET/CT imaging.

CNN is a type of deep learning method that constitutes two parts: back-propagation and optimization. The loss function is used to measure the distance from the desired solution. In the classification of datasets, the desired loss of the function used in back-propagation is cross-entropy [20, 21]. The in-built CNN designer in MatLab 2020b was used in our study, and a CNN with 40 hidden neurons and two softmax output neurons was created. As mentioned earlier, the training, validation, and testing sets were assigned. Supervised training is when the results are fed to CNN, and the efficiency of the CNN is then calculated [20]. It provides a perceptive result. Supervised training was performed. In our study, CE was 13.6% in classifying csPCa and insignificant PCa.

PET/CT is one of the high-end investigations available in oncology, and it can provide valuable information about the behavior of cancer. It can guide the development of better screening tools, treatment planning, prognostication, and response evaluation [22]. ⁶⁸ Ga PSMA PET/CT is the currently preferred modality for detecting metastasis during primary prostate cancer initial imaging [4, 5]. ⁶⁸ Ga PSMA PET/CT can also detect index lesions in the case of multifocal lesions and high-risk lesions because high ⁶⁸ Ga PSMA uptake is seen in aggressive lesions and lesions with high Gleason scores [23–25]. PSMA PET/CT has shown a higher detection rate of csPCa in patients with PSA levels 4–20 ng/ml [5].

PET/CT-guided biopsy is an upcoming approach in clinical oncology and has increased diagnostic accuracy to establish a pathological diagnosis over conventional image-guided sampling [26, 27]. PSMA PET/CT-guided prostate biopsy has higher diagnostic accuracy and lesser complication rates than TRUS-guided biopsy [6]. ⁶⁸ Ga PSMA PET/CT-guided biopsy is being done routinely in our center for initial diagnosis, recurrence evaluation, and metastasis evaluation. Since ⁶⁸ Ga PSMA PET/CT has many advantages over MRI in providing more reliable histological characteristics due to ⁶⁸ Ga PSMA uptake by the tumor cells, ⁶⁸ Ga PSMA PET/CT radiomic features are exceptionally steadfast. Additionally, the high diagnostic accuracy of ⁶⁸ Ga PSMA PET/CT-guided prostate biopsy results in precise histological results, making the histological reports unquestionable. These features of this study make the supervised CNN training well-grounded. The present study demonstrated that extracted RF in these patients had shown higher accuracy for differentiating clinically significant and clinically insignificant prostate cancer.

One limitation of this study is that inter-observer variability can occur during ROI creation, which was not considered in the present study and was reported previously [28]. Secondly, the sample size is lesser compared to other radiomics studies [8]. Automated processing of the images to extract radiomic features was not performed due to technical difficulties and expert nuclear medicine physician can avoid potential pitfalls common in ⁶⁸ Ga PSMA PET/CT which can be missed by the machine learning processing.

Conclusion

Deep learning is a feasible pre-biopsy screening tool for identifying significant prostate cancer. It can be used as an adjunct in the initial diagnosis and for early treatment planning. We also aim to further validate this study in the future by recruiting more patients and by comparing it with other machine learning methods with higher computational capacity. Further studies are needed to increase data availability and improve the performance of cheap AI tools.

Supplementary Information

Below is the link to the electronic supplementary material.

Author Contribution

All authors have made a significant contribution in formulating, reviewing, and editing the manuscript.

Data Availability

Data generated and analyzed can be availed from the corresponding author on a reasonable request.

Declarations

Competing Interests

Rajender Kumar, Arivan Ramachandran, Bhagwant Rai Mittal and Harmandeep Singh declare no competing interests.

Ethics Approval and Consent to Participate

This study was performed in accordance with the ethical standards as laid down in the 2013 Declaration of Helsinki and its later amendments or comparable ethical standards. The institutional review board waived the requirement for written consent.

Consent for Publication

Informed consent was obtained from the participant for the study mentioned as a part of the institutional protocol.