Novel Bone Scan Features for Predicting Prognosis in Men With Bone Metastatic Prostate Cancer: A Retrospective Study

Abstract

Background

Bone metastasis frequently occurs in patients with prostate cancer, however, a consensus has not been reached regarding bone scan image analysis. We aimed to analyse various bone scan imaging features of metastatic prostate cancer and to assess their impact on prognosis.

Methods

One thousand five hundred sixty-three paired sets of bone scan images (anterior and posterior) were obtained from patients with metastatic prostate cancer at Seoul National University Hospital. U-Net architecture was used for the segmentation of metastatic bone lesions. Imaging features describing the overall metastatic burden (n = 18) and largest metastatic burden (n = 32) were extracted using computer vision techniques. Kaplan-Meier survival analysis and Cox proportional risk model were used to analyse the prognostic impact of each feature.

Results

The correlation coefficient between the actual number of lesions and that predicted by the deep learning model was 0.87, indicating a strong correlation. Multivariate Cox regression showed that metastasis intensity difference (hazard ratio [HR], 0.53; P = 0.002) and the largest metastasis percentage (HR, 0.62; P = 0.038) were independently associated with disease progression and were even more strongly associated with the number of metastases (current standard). The Kaplan-Meier curves revealed that a higher total metastasis ratio (P < 0.001), a lower total metastasis intensity difference (P = 0.030), a lower largest metastatic lesion percentage (P < 0.001), higher compactness (P = 0.028), and lower eccentricity (P = 0.070) were associated with shorter progression-free survival.

Conclusion

Although the number of bone metastases is a standardised prognostic factor, additional consideration of morphological or intensity-related novel features may be useful to more accurately predict the prognosis of patients with metastatic prostate cancer.

Graphical Abstract

INTRODUCTION

Due to bone tropism, approximately 70–80% of patients with metastatic prostate cancer (PCa) present with or develop bone metastasis.1 Bone scintigraphy (bone scan) has emerged as a vital diagnostic tool for detecting and monitoring advanced PCa.2 Accurate assessment of the patient’s bone metastatic status through bone scan imaging is mandatory for proper treatment decisions and treatment response validation.

In previous randomised clinical trials (RCTs) (CHAARTED and LATITUDE), bone metastasis status was conventionally divided into high and low burden according to the number of metastases, which was associated with different overall survival.3,4 However, accumulating evidence has shown that the number of metastases alone cannot fully elucidate a patient’s metastatic status. For example, in a recent phase 3 RCT by Armstrong et al.,5 automated Bone Scan Index (BSI), which is a quantitative measure of the percentage of the adult skeleton involved in bone metastasis, was an independent prognostic imaging biomarker of survival in metastatic castration-resistant PCa.6 Moreover, Roth et al.7 recently demonstrated the prognostic value of not only the whole fraction of bone metastasis but also the sub-fractions according to each anatomical location, as well as the intensity of uptake. They concluded that the quantitative characterisation of metastatic spread by anatomic location enhances potential prognostication.

In this study, we propose new quantitative features related to prognostic prediction in patients with bone metastatic PCa that were not utilized in prior studies. These new features were extracted using deep learning, leveraging 50 morphological and intensity characteristics to reflect differences in the spatial and intensity heterogeneity of bone metastases. Through survival analysis, we validated these features as potential biomarkers for more precise prognostication of bone metastatic PCa, offering an automated and reproducible approach.

METHODS

Study design and population

We retrospectively analysed 207 patients who were initially diagnosed with bone metastatic PCa at Seoul National University Hospital between March 2000 and December 2019. All patients were initially treated with androgen deprivation therapy, and whole-body positron emission tomography (PET)-magnetic resonance imaging or bone scan tests were performed at the time of diagnosis. We collected 1,563 paired sets (anterior and posterior) of bone scan image datasets, and bone metastasis lesions were carefully annotated by nuclear medicine and urology experts based on interpretation reports, ensuring meticulous exclusion of false-positive findings, such as arthritis and traumatic changes.

Data preprocessing

Because of the different intensity ranges and sizes of the Digital Imaging and Communications in Medicine (DICOM) images, pre-processing was performed as follows. Here, intensity refers to the pixel value in the DICOM images, representing the grayscale level. First, the maximum intensity value was adjusted based on the pixel value count with high intensity, and pixel values were rescaled to a range of 0–255. Second, all images were centrally cropped and resized to 256 × 1,024 pixels (width by height), and we reversed the left and right ground truth (mask images) of posterior bone metastasis and then combined the anterior ground truth and posterior ground truth into one image.

Metastatic bone lesion segmentation model

We used the U-net8 model, which performed well in medical image segmentation for bone metastasis segmentation. The input consisted of two-channel images, with anterior and posterior bone scan images for each channel. Segmentation was performed simultaneously by different networks, and each result was concatenated in the last layer to produce a segmentation map. For all inputs, we assessed a Dice coefficient between the final resulting image of the model and the ground truth of bone metastasis obtained by merging the anterior and posterior images (Fig. 1). The dataset comprised 1,100, 232, and 231 images from the training, validation, and test sets, respectively. The activation function of the model used stochastic gradient descents with a momentum of 0.9 to optimise the overall parameters. We trained the model with a learning rate of 1e-2, step decay of 1e-6, and batch size of 4.

Fig. 1. Overall model pipeline of bone metastasis segmentation. Segmentation is performed on both anteroposterior and posterior-anterior images. Then, intensity and morphological features of the metastatic bone lesions are automatically extracted.

Feature extraction of metastatic bone lesions

For the largest lesion and all metastatic lesions, we extracted intensity and morphological features from the segmented bone metastasis maps using a deep learning model. To extract intensity features, we first detected the overlapping regions of the bone scan and segmented bone metastasis maps. Next, the actual intensities of the detected regions were measured. Using the measured values, we quantified the intensity features, such as the maximum, minimum, and standard deviation. Additionally, for the largest lesions, we quantified the intensity difference between the largest lesion and the surrounding areas. For morphological features, computer vision algorithm was performed. Density was quantified by calculating the distance between detected lesions. A percentage of metastasis was measured by calculating the area of the whole-body bone and the size of the area of bone metastasis. The size of the overall metastasis and size of the largest metastasis were measured, the proportion occupied by the largest metastatic lesion was quantified. For the largest lesion, the degree of irregularity in the shape of the lesion was measured using the following equation.

We used compactness as in equation9:

$$Compactness = \frac{{\left(Perimeter\right)}^2}{4\pi \times Area} \left(1\right)$$

$$Eccentricity = \frac{{Axis Length}_{Short}}{{Axis Length}_{Long}} \left(2\right)$$

Statistical analysis

For each imaging feature, patients were divided into two (high vs. low) based on the quantified median value of that feature. The median value for each variable was calculated separately, and this value was used as the threshold to classify patients into high and low groups for survival analysis. Univariate and multivariate Cox regression analyses were performed to assess the independent effect of possible prognostic factors on progression-free survival (PFS). All statistical analyses were performed using SPSS software (version 25.0.0.2; IBM Corp., Armonk, NY, USA). Statistical significance was set at P < 0.05, and all statistical tests were two-sided.

Ethics statement

This study was approved by the Institutional Review Board (IRB) of Seoul National University Hospital (IRB number: 2008-081-1148) and performed in accordance with the applicable laws and regulations, good clinical practice, and ethical principles as described in the Declaration of Helsinki. The requirement for informed consent was waived owing to the retrospective study design.

RESULTS

Metastatic PCa study cohort

This study enrolled 207 metastatic PCa patients with a mean age of 70.9 ± 8.9 years. Among these patients, 85 (41%) had 10 or fewer metastatic lesions, whereas 122 (59%) had more than 10 metastatic lesions. Most patients (n = 175, 84.5%) were categorised in the International Society of Urological Pathology Grade Group (ISUP GG) 4–5. There were 141 (68.1%) patients with lymph node metastases and 35 (16.9%) with visceral metastases (Table 1).

Table 1. Baseline characteristics of metastatic prostate cancer patients whose images were used for the survival analysis.

Variables		Values (N = 207)
Age, yr		70.9 ± 8.9
With HTN		71 (34.3)
With DM		44 (21.3)
No. of bone metastasis
	≤ 10	85 (41.1)
	> 10	122 (58.9)
T stage
	T2	57 (27.5)
	T3–4	175 (84.5)
ISUP GG
	GG1–3	32 (15.5)
	GG4–5	175 (84.5)
PSA at ADT initiation		278 (101–657)
LN metastasis		141 (68.1)
Visceral metastasis		35 (16.9)

Values are presented as number (%), mean ± standard deviation, or median (interquartile range).

HTN = hypertension, DM = diabetes mellitus, ISUP GG = International Society of Urological Pathology Grade Group, PSA = prostate-specific antigen, ADT = androgen deprivation therapy, LN = lymph node.

Segmentation performance of metastatic bone lesions

The performance of the segmentation model was evaluated as 0.52 according to the Dice coefficient score. Supplementary Fig. 1 shows a strong correlation10 (r = 0.87; P < 0.001) between the actual number of lesions and the predicted number of lesions by the deep learning model in the segmentation map.

Metastatic bone lesion features

We extracted a total of 50 characteristic features of bone metastasis from a segmented bone metastasis map using advanced computer vision techniques. On the basis of the intensity distribution, the features related to the intensity of the 36 lesions were extracted (Fig. 2). The intensity feature includes features such as the minimum and maximum values of the intensity of the overall lesion, difference between the minimum and maximum values, and difference between the lesion and the surrounding area intensity. The remaining 14 morphological features were extracted using shape analysis.9 Morphological features included size, perimeter, compactness, and eccentricity of the largest lesion among the detected lesions (Fig. 3). The more irregular the shape of the lesion, the greater the value of compactness; however, the more regular the shape of the lesion, the greater the value of eccentricity.

Fig. 2. Intensity features. (A) Bone scan; (B) Segmentation model output; (C) Bone metastasis detection; (D) Intensity distribution.

Fig. 3. Representative demonstration of a morphological feature and its value. (A) Eccentricity; (B) Solidity; (C) Compactness.

Prognostic assessment

For each imaging feature, the patients were divided into high and low groups. In univariate Cox regression analysis, a large number of bone metastases (hazard ratio [HR], 2.07; P < 0.001) were associated with short PFS, whereas high T stage (P = 0.520), high ISUP GG 4–5 (P = 0.711), existence of lymph node metastasis (P = 0.534), or visceral metastasis (P = 0.253) were not associated with short PFS. Regarding bone scan imaging features, the total metastasis ratio (metastatic bone area/total bone area) (HR, 1.98; P < 0.001), metastasis intensity difference (maximum intensity − minimum intensity) (HR, 0.66; P = 0.038), and standard deviation of metastasis intensity (HR, 0.71; P = 0.080) were significant prognostic bone scan imaging features for the overall assessment. The largest metastatic lesion feature assessment included the largest metastatic percentage (largest metastatic bone area/total metastatic area) (HR, 0.47; P < 0.001), largest metastasis compactness (HR, 1.53; P = 0.031), eccentricity (HR, 0.71; P = 0.078), and standard deviation of boundary (HR, 0.57; P = 0.005) (Table 2). In multivariate Cox regression analysis, the overall metastasis intensity difference (HR, 0.53; P = 0.002), largest metastasis percentage (HR, 0.62; P = 0.038), and standard deviation of the largest metastasis boundary (HR, 0.68; P = 0.053) were associated with PFS, independent of the number of bone metastases (> 10) (Table 3).

Table 2. Univariate Cox regression analysis for progression-free survival in metastatic prostate cancer patients.

Variables			HR (95% CI)	P value
Age, yr			1.00 (0.98–1.03)	0.787
T stage 3–4			0.88 (0.59–1.31)	0.520
ISUP GG 4–5			1.10 (0.67–1.81)	0.711
PSA			1.00 (1.00–1.00)	0.479
LN metastasis			0.88 (0.59–1.31)	0.534
Visceral metastasis			1.34 (0.81–2.20)	0.253
Bone metastasis > 10			2.07*** (1.38–3.12)	< 0.001
Bone metastasis AI featuresa
	Overall featurea
		Total metastasis ratio (metastatic bone/total bone)	1.98*** (1.35–2.92)	< 0.001
		Metastasis intensity difference (maximum intensity − minimum intensity)	0.66 (0.45–0.98)	0.038
		Metastasis intensity standard deviation	0.71 (0.48–1.04)	0.080
	Largest metastatic lesion featurea
		Largest metastasis percentage (largest metastatic bone/total metastasis)	0.47*** (0.32–0.70)	< 0.001
		Largest metastasis compactness (irregularity)	1.53 (1.04–2.24)	0.031
		Largest metastasis eccentricity	0.71 (0.48–1.04)	0.078
		Largest metastasis boundary standard deviation	0.57 (0.39–0.84)	0.005

HR = hazard ratio, CI = confidence interval, ISUP GG = International Society of Urological Pathology Grade Group, PSA = prostate-specific antigen, LN = lymph node, AI = artificial intelligence.

^aAI features were applied as binary variable (high vs low).

^***P < 0.001.

Table 3. Multivariate Cox regression analysis for progression free survival.

Variables			HR (95% CI)	P value
Model 1
	Bone metastasis > 10		1.65 (0.91–2.97)	0.097
	Overall bone metastasis featurea
		Total metastasis ratio (metastatic bone/total bone)	1.65 (0.94–2.90)	0.084
		Metastasis intensity difference (maximum − minimum)	0.53*** (0.36–0.79)	0.002
Model 2
	Bone metastasis > 10		1.44 (0.88–2.35)	0.143
	Largest metastatic lesion featurea
		Largest metastasis percentage (largest metastatic bone/total metastasis)	0.62 (0.40–0.97)	0.038
		Largest metastasis compactness (irregularity)	1.17 (0.76–1.81)	0.484
		Largest metastasis eccentricity	1.16 (0.76–1.76)	0.498
		Largest metastasis boundary standard deviation	0.68 (0.45–1.00)	0.053

HR = hazard ratio, CI = confidence interval.

^aAI features were applied as binary variable (high vs low).

^***P < 0.001.

Survival curve analysis

Overall lesions, a high total metastasis ratio (median PFS, 29 vs. 56 months; log-rank, P < 0.001) and low total metastasis intensity difference (median PFS, 27 vs. 44 months; log-rank, P = 0.030) were associated with short PFS. For the largest metastatic lesion feature, those without the largest dominant metastatic lesion (median PFS, 25 vs. 56 months; log-rank, P < 0.001), a high compactness score (the degree of irregularity) (median PFS, 32 vs. 51 months; log-rank, P = 0.028), a low eccentricity score (ellipsoid rather than round) (median PFS, 31 vs. 42 months; log-rank, P = 0.070), and a low standard deviation of boundary intensity (median PFS, 27 vs. 47 months; P = 0.004) were associated with short PFS (Fig. 4).

Fig. 4. Kaplan-Meier curve of the representative bone scan imaging features. (A) Total metastasis ratio; (B) Largest metastasis percentage; (C) Largest metastasis compactness; (D) Largest metastasis eccentricity.

DISCUSSION

In this study, metastatic PCa was segmented through deep learning, and from the segmented maps, morphological and intensity-related imaging features were extracted using computer vision algorithms. Through Cox regression and Kaplan-Meier curve analyses, we identified new biomarkers that can predict the prognosis of patients with metastatic PCa, surpassing conventionally used features, such as the number of lesions or BSI. These biomarkers may offer a more biologically nuanced stratification of risk among patients with metastatic PCa, bringing us closer to personalised treatment approaches.

Recently, using artificial intelligence, several studies have attempted to effectively and rapidly assess the bone metastatic burden using bone scan images,2,11,12 however, research to find new imaging biomarkers has rarely been conducted. In this context, our novel findings may be helpful in assessing early progression and performing proper early intervention, which may lead to better oncological outcomes and improved bone health with fewer skeletal-related events.13

According to our study, several imaging features were more prominently associated with survival outcomes than already known features, such as the number of bone metastases (conventional criteria) and relative uptake area (BSI). For the number of bone metastases, we found that the conventional number suggested by authors of the CHAARTED trial (≥ 4) is not appropriate for our study population (Supplementary Fig. 2), which is in line with findings of the previous study by Yamada et al.14 They demonstrated that the CHAARTED and LATITUDE criteria of ‘high-volume’ or ‘high-risk’ are not appropriate for Asian patients.14 They insisted that ≥ 11 bone metastases with the highest HR may be a suitable definition for ‘high-volume’ PCa in Asians. The result has been reproduced in our cohort, showing that a cut-off value of 11 is better than 3 or 4 for the number of metastatic lesions; thus, number of metastatic lesion of ≥ 11 was used as a prognostic factor and used for survival analysis.

In the model using the overall bone metastasis assessment features, some imaging features had comparable or better prognostic effects than the number of metastases. The metastasis intensity difference had a better prognostic impact than BSI or the number of bone metastases (> 10). Differences in metastatic intensity indicate the degree of intensity range or heterogeneity. Several theories have supported this hypothesis. First, different intensities may indicate false-positive bone lesions,15,16 indicating the possibility of an overrepresentation of the metastatic burden status. Second, heterogeneous uptake (intensity) may indicate rapid turnover of metastatic bone lesions with actively newly developed lesions, which may be better treated than already-seeded stable lesions. A similar feature on PET/computed tomography has been suggested as a possible prognostic factor for breast cancer.17 In addition, Francis et al.18 reported that, based on a study of primary bone tumours, various radionuclide uptake levels may be associated with blood flow and the degree of osteoblastic activity, which is also related to the patient’s response to therapy. Collectively, disease status with a high-intensity difference may indicate a good treatment response, leading to longer PFS.

In the model using the largest metastatic bone lesion or index lesion features, we suggest some novel prognostic findings. This analysis was performed on the basis of the hypothesis that the prognosis is largely determined by the main lesion (the largest metastatic lesion). The largest metastasis percentage and standard deviation of boundary intensity were independently prognostic from the number of metastases. This indicates that patients with one dependent metastatic lesion and smaller additional lesions showed more favourable outcomes than those without representative index lesions, if the patients had the same number of metastatic lesions. Multiple lesions of similar size may indicate similar active but heterogeneous metastatic clones, suggesting the possibility of different treatment responses, leading to a short PFS. For boundary intensity, a high standard deviation indicated that the tumour border-aligned cancer cells had varying intensities. This may mean that the lesions have recently developed and actively interacted with the surrounding tissue, which may benefit from therapeutic intervention that aims to disrupt the interaction, leading to a good prognosis from early intervention.19

Several interesting features were associated with PFS in Kaplan-Meier curve analysis. Among these, the shape of metastatic lesions is interesting. We demonstrated that two imaging features, i.e. compactness and eccentricity, were associated with PFS. If the metastatic lesion contour was regular rather than irregular and sharp and if the lesion was round rather than ovoid, the patient had better survival (Fig. 4). This is in line with a previous finding that metastatic bone lesions with speculated periosteal reactions indicate rapidly progressive disease in metastatic PCa.20

This study has several limitations. First, because the pixel range of the collected bone scan image data is not the same, we proceeded with the process of normalising the pixels, and the data of the image were lost during the pre-processing process; thus, the Dice coefficient score seems to be low. Second, although we discovered novel imaging features related to PFS, this is still based on a hypothetical theory and the demonstration of association rather than causal relationships affecting survival. Mechanistic or prospective observational studies are warranted to further investigate each imaging feature during the treatment course. Third, the study could not investigate the clinical significance of changes in the novel features observed in bone scan images during the treatment process, as we only used bone scan images obtained at the initial diagnostic stage. Fourth, external validation is required with a large patient population. Therefore, we plan to conduct a multicentre study to validate our findings. Finally, although a bone scan is a cost-effective and sensitive test for detecting metastatic disease, it lacks specificity. Thus, not every lesion may be a true metastasis, which should be confirmed using hybrid imaging.16

SUPPLEMENTARY MATERIALS

Supplementary Fig. 1

Confusion matrix showing the correlation between the number of bone metastases defined by the model and the actual number calculated by the expert.

Supplementary Fig. 2

Kaplan-Meier curve of the effect of the number of bone metastases on prognosis. Number of bone metastases: (A) > 10 and ≤ 10; (B) > 3 and ≤ 3.