Computer-assisted scatter plot analysis of cell and nuclear areas distinguishes urothelial carcinoma in urine cytology specimens

INTRODUCTION

Urine cytology is a simple, non-invasive test commonly used for the early detection of urothelial malignancies and for monitoring patients following urothelial carcinoma surgery.^[1,2] While urine cytology typically has high specificity, its sensitivity is often low.^[3] Reported sensitivity rates are 77% for muscle-invasive urothelial carcinoma, 59% for high-grade non-muscle-invasive urothelial carcinoma, and 33% for low-grade non-muscle-invasive urothelial carcinoma.^[4] For noninvasive cancers, especially pTa lesions, cytology sensitivity remains low.^[5] As a result, despite significant differences in treatment approaches between non-invasive and invasive carcinomas,^[2] their diagnosis currently relies on histological evaluation.

In recent years, image analysis and machine learning have been increasingly used to differentiate normal urothelial cells from high-grade urothelial carcinoma (HGUC)^[6] and to identify high-grade and atypical urothelial cells in urine cytology specimens.^[7] These methods have also been applied to assess the risk of urothelial carcinoma recurrence based on nuclear morphology in histological specimens^[8] and to predict recurrence in non-muscle-invasive urothelial carcinoma by combining clinical and histological data.^[9] Many studies have emphasized nuclear morphology, particularly nuclear area, perimeter, maximum diameter, and minimum diameter. Sakumo et al. reported that the nuclear area and the nuclear-cytoplasmic (N/C) ratio could help distinguish reactive urothelial cells and reactive tubular epithelial cells from HGUC.^[10] Yamamoto et al. examined non-invasive follicular thyroid neoplasm with papillary-like nuclear features (NIFTP), follicular carcinoma, and papillary thyroid carcinoma follicular subtype using parameters such as nuclear area and perimeter.^[11] Unlike image analysis, routine cytological evaluation by human experts also considers cell clusters for cellular overlap, density, and loose cell adhesion, in addition to nuclear morphology.^[12,13] There is a disparity between the indices used in image analysis and those employed by human evaluators. Only one study by Levy et al. has focused on analyzing cell clusters in cytological specimens, where they developed an automated tool for evaluating urothelial cell clusters.^[14]

In this study, we explored the potential of incorporating “overlapping of cell clusters” and “cell density within clusters” into the image analysis algorithm. Using urine cytology, we visualized these features as scatter plots showing the area of cell clusters and the maximum connected nuclear area within them.

MATERIAL AND METHODS

RESULTS

Scatter plots of the morphology of the nuclear areas and nuclear perimeters were ineffective in distinguishing tumor samples from non-tumor samples in urinary cytology specimens

Table 1 presents the clinicopathological characteristics of the cases included in this study, while Table 2 details the clinical main complaints or diagnoses for the cytology-negative cases. The scatter plots for the nuclear area and nuclear perimeter were generated because these metrics are commonly used in the image analysis of cytological specimens.^[10,16] The scatter plot for the nucleus was compared between the cytology-negative and cytology-positive groups [Figure 1]. The percentage of nuclei in the tumor-rich areas was significantly higher in the cytology-positive group compared to the negative group (P = 0.0134). When the positive group was divided into non-invasive and invasive tumor groups, a significant difference was observed between the non-tumor and noninvasive tumor groups (P = 0.0281). However, no significant difference was found between the non-tumor and invasive tumor groups (P = 0.1266). These findings indicate that while it is possible to distinguish between the cytology-negative and cytology-positive groups using the scatter plot for the nucleus, this method is not robust, as it fails to distinguish between the invasive tumor group and the non-tumor group.

Table 1: Clinicopathological characteristics of the cases included in this study.

Cases (n=192)				Cytological diagnosis
			Negative	Suspicious	Positive
	Negative	52	52 (100)
Stage
Muscle invasion (−)	pTa (low grade)	4		1 (25)	3 (75)
	pTa (high grade)	15		6 (40)	9 (60)
	pTis	21		7 (33)	14 (67)
	pTa+pTis	15		6 (40)	9 (60)
	pT1	41		10 (24)	31 (76)
Muscle invasion (+)	pT2	31		10 (32)	21 (68)
	pT3	9		3 (33)	6 (67)
	pT4	4		1 (25)	3 (75)

The numbers outside the brackets indicate the number of cases, and the numbers inside the brackets indicate the proportion of cases in the row direction

Table 2: Main clinical complaints or diagnoses in the participants with negative cytology results for urothelial carcinoma.

Main clinical complaints or diagnoses	Breakdown
No tumor history	33
Hematuria alone	14
Hydronephrosis	5
After kidney transplantation	4
Prostatic hyperplasia+hematuria	3
Overactive bladder	1
Essential thrombocytosis+hematuria	1
Frequent urination	1
Hemorrhagic cystitis+hematuria	1
Pyuria+overactive bladder	1
Chronic renal failure+pyonephrosis	1
Amyloidosis+hematuria	1
Tumor history	19
Suspected bladder cancer	6
Prostate cancer follow-up	5
Suspicious of ureteric cancer	1
Bladder cancer follow-up	1
Suspected renal pelvis cancer	1
Renal mass	1
Breast cancer, treatment	1
Stomach cancer, post-operative	1
Uterine cancer, post-operative	1
Cervical cancer	1

Close

Figure 1:

Scatter plot representing the nuclear area and perimeter of the cells, along with a comparison of the positive rates for high nuclear area and perimeter in the total cell nuclei. (a) Scatter plot for a cytology-negative case (Case 43). (b) Scatter plot for a cytology-positive case (Case 190). (c) Comparison of the positive rates of high nuclear area and perimeter in total cell nuclei between cytology-negative and cytology-positive cases. (d) Comparison of the positive rates of high nuclear area and perimeter in total cell nuclei across cytology-negative cases, invasion-negative tumor cases, and invasion-positive tumor cases. The “X” marks the mean values in the box plot, while boxes with horizontal lines represent the medians with interquartile ranges. Error bars indicate maximum and minimum values, and circles represent outliers. P-values were determined using Dunnett’s analysis and the Tukey–Kramer test. ^✶P < 0.05. Thus, the largest nuclear area/cellular area in b) is 9.22%, and the one in d) is 36.71%.

Export to PPT

Scatter plots combining cells/cell cluster area and maximum nuclear area revealed distinct pattern differences between the cytology-negative and cytology-positive groups

The scatter plots for cell/cell clusters are presented in Figures 2 and 3. In the representative cell clusters of the cytology-negative group, the nuclei did not overlap, resulting in an occupancy rate of the maximum nuclear area within the cell/cell cluster of 9.22% [Figure 2a and b]. In contrast, the cytology-positive group exhibited high cellular density and a high N/C ratio, causing the nuclear area of individual tumor cells to be recognized as a connected nuclear area within the tumor cell cluster. Consequently, the occupancy rate of the maximum nuclear area within the tumor cell aggregate was 36.71% [Figure 2c and d]. For the scatter plot of cell area versus maximum nuclear area in the cytology-negative group, the plots were distributed along the X-axis (horizontal pattern). In the cytology-positive group, however, the plots were distributed along the Y-axis (vertical pattern). In addition, both groups displayed other plotting patterns, where the plots were distributed diagonally (diagonal pattern) [Figure 3]. A stacked bar graph illustrated the ratios of horizontal, vertical, and diagonal patterns among the non-tumor and malignancy groups [Figure 4a]. In the non-tumor group, the predominant plotting pattern was horizontal (57.7%), while the malignancy group primarily exhibited a vertical pattern (73.6%). A significant difference was found between the groups (P < 0.0001, Pearson’s Chi-square value: 104.481). The ratios of scatter plot patterns were then compared across the non-tumor, non-invasive, and invasive tumor groups [Figure 4b]. The non-tumor group mainly displayed a horizontal pattern (57.7%), whereas the noninvasive and invasive tumor groups predominantly showed vertical patterns (78.2% and 70.6%, respectively). Pearson’s Chi-square test indicated a significant difference among the three groups (P < 0.0001, and Pearson’s Chi-square value: 105.350). Moreover, differences in plot patterns within the non-invasive tumor group were examined. As shown in Figure 4c, the horizontal pattern was most common in the non-tumor group (57.7%), while the diagonal pattern was mainly observed in the pTa low-grade group (75%). The vertical pattern was predominantly seen in the pTa high-grade group, pTis group, and pTis + pTa group (68.8%, 81.0%, and 100%, respectively). This suggests that, despite differences in grade or invasion depth, the tumor groups exhibited distinct scatter plot patterns compared to the non-tumor group. Therefore, the scatter plot for cell/cell clusters could be useful not only in distinguishing the cytology-negative group from suspicious malignancy or malignancy groups but also in distinguishing early-stage cancer from the non-tumor group.

Close

Figure 2:

Example of computer-assisted image analysis for normal urothelial cells and high-grade urothelial carcinoma (HGUC) cell clusters. (a and b) Normal urothelial cell cluster (Case 17). (c and d) HGUC case (Case 77). (a and c) show images before analysis, while (b and d) display images after analysis. The cellular area of the normal urothelial cell cluster in the analyzed image is outlined by the light green line, and the individually detectable nuclear areas are framed by the light blue line (b). The light green line outlines the cellular area of the carcinoma cell cluster in the analyzed image, and the individually detectable nuclear-connected area is framed by the light blue line (d). The actual cellular area data in (b and d) are highlighted with a green background and green line, while the largest nuclear area data in the cell cluster in (b and d) are highlighted with a red background and red arrow. The second-largest nuclear area data in the cell cluster in (b and d) are highlighted with a blue background and blue line. Thus, the largest nuclear area/cellular area in b) is 9.22%, and the one in d) is 36.71%.

Export to PPT

Close

Figure 3:

Plot of the cell and cell aggregate areas alongside the largest nuclear area within the cell/cell aggregate. (a) Horizontal pattern observed in the cytology-negative case (Case 34) and (b) vertical pattern observed in the cytology-positive case (Case 190). (c) Diagonal pattern in the cytology-negative case (Case 1) and (d) diagonal pattern in the cytology-positive case (Case 181).

Export to PPT

Close

Figure 4:

A stacked bar chart illustrating the composition rates of each scatter plot pattern for the cell/cell aggregate areas and their largest independently detectable nuclear areas. (a) Comparison of scatter plot patterns based on cytological diagnosis. (b) Comparison of scatter plot patterns among the non-tumor, non-invasive, and invasive tumor cases. (c) Comparison of scatter plot patterns between non-tumor cases and non-invasive tumor categories. ^✶✶P < 0.0001.

Export to PPT

The percentage of cells/cell clusters in tumor-rich areas was significantly higher in the tumor group than the non-tumor group, but it was lower in the invasive tumor group than in the non-invasive tumor group

Since the predominant plotting pattern of the scatter plot for the tumor group was vertical, it was considered that tumor cells or cell clusters might be found in areas with a cell/cell cluster area of <2000 μm² and a maximum nuclear area >75 μm² (tumor-rich area), as cell/cell clusters were rarely observed in the scatter plot of the non-tumor group. Thus, we calculated the percentage of cell/cell clusters in the tumor-rich area [Figure 5a and b]. To verify the presence of tumor cells in this area, we examined whether tumor cells or tumor cell clusters were present in the tumor-rich area. As shown in Figure 5c, we confirmed that tumor cell clusters consisting of 10–20 tumor cells were located in the upper part of the tumor-rich area along the Y-axis [Figure 5c (a-c)]. In contrast, single tumor cells were found in the lower part of the tumor-rich area along the Y-axis [Figure 5c (d-f)]. This suggests that the area defined for the tumor cell or cell cluster-rich area was appropriately set. The percentage of cell/cell clusters in the tumor-rich area was analyzed for all samples [Figure 6]. The percentage of cell/cell clusters in the tumor-rich area for both the non-invasive tumor group (pTa, pTis) and the invasive tumor group (pT1–pT4) was higher than that of the non-tumor group. However, the percentage of cell/cell clusters in the tumor-rich area was significantly lower in the invasive tumor group compared to the noninvasive tumor group (P = 0.0299) [Figure 6a]. Moreover, the percentage of cell/cell clusters in the tumor-rich areas for both the non-muscle-invasive and muscle-invasive groups was greater than that of the non-tumor groups. While the muscle-invasive tumor group did not show statistically significant differences compared to the non-muscle-invasive tumor group, there was a tendency toward a difference (P = 0.0574) [Figure 6b]. These results suggest that the tumor group can be distinguished from the non-tumor group based on the percentage of cell/cell clusters in the tumor-rich areas, regardless of the depth of tumor invasion. However, the lower percentage of cell/cell clusters in the tumor-rich areas of the invasive tumor and muscle-invasive tumor groups indicates an underlying reason for this observation.

Close

Figure 5:

Plot of the cell/cell aggregate area and their largest nuclear area within the cell/cell aggregate, alongside actual cell image in the tumor-dominant area. (a) Scatter plot for the cytology-negative case (Case 34). (b) Scatter plot for the cytology-positive case (Case 190). The thick black square lines in both (a and b) indicate the tumor-dominant area. (c) Magnified scatter plot (Case 111) showing the distribution of individual tumor cells and small to medium-sized tumor cell clusters in the tumor-dominant area. The red line denotes the area of small to medium-sized tumor cell clusters and representative images of those clusters, respectively. The blue line indicates areas of individual tumor cells and representative images of those cells. Each cell or cell cluster area and maximum nuclear area shown in Figure 5c (a-f) were as follows: 1507.77 μm² and 465.60 μm² (a), 1625.28 μm² and 993.72 μm² (b), 1390.41 μm² and 776.52 μm² (c), 155.85 μm² and 113.65 μm² (d), 243.18 μm² and 158.49 μm² (e), and 152.19 μm² and 118.98 μm² (f), respectively.

Export to PPT

Close

Figure 6:

Comparison of the positive rates for the largest nuclear area within a high cell/cell aggregate-distributed area. (a) Comparison among non-tumor, non-invasive, and invasive tumor cases. (b) Comparison between non-tumor cases, non-muscle-invasive tumor cases, and muscle-invasive tumor cases. “X” marks the mean values in the box plot, while boxes with horizontal lines indicate the medians with interquartile ranges. Error bars represent the maximum and minimum values, and circles denote outliers. P-values were calculated using the Tukey–Kramer test. ^✶P < 0.05 and ^✶✶P < 0.0001.

Export to PPT

Neutrophil exudation was higher in the invasive tumor group compared to the non-invasive tumor group, and it was also greater in the muscle-invasive tumor group than in the non-muscle-invasive tumor group

We examined the virtual slide images of all cases for neutrophil exudation, with representative images of poor and rich neutrophil exudation, as shown in Figure 7. When comparing neutrophil exudation between the non-invasive tumor group (pTa and pTis) and the invasive tumor group (pT1–pT4), the background neutrophil exudation was significantly higher in the invasive tumor group (P = 0.0191, Pearson’s Chi-square value: 5.490) [Table 3]. Similarly, when comparing the non-muscle-invasive tumor group (pTa, pTis, and pT1) with the muscle-invasive tumor group (pT2–pT4), the background neutrophil exudation was significantly higher in the muscle-invasive tumor group (P = 0.0014, Pearson’s Chi-square value: 10.269) [Table 4]. These results suggest that the lower percentage of cell/cell clusters in tumor-rich areas of the invasive and muscle-invasive tumor groups, compared to the non-invasive and non-muscle-invasive group, may be attributed to neutrophil exudation into the urine.

Close

Figure 7:

Comparison of background neutrophil infiltration levels and tumor invasion depth. Left: Example of low background neutrophil infiltration (Case 123). Right: Example of high background neutrophil infiltration (Case 187).

Export to PPT

Table 3: Confusion matrix showing the relationship between background neutrophil infiltration and tumor invasion depth (non-invasive versus invasive) in urothelial carcinoma patients.

Depth of invasion	Background neutrophil infiltration			Chi-square
	Low	High	P-value
Non-invasive (pTa, pTis)	44	11	0.0191✶	5.490
Invasive (pT1-pT4)	52	33

P-value was calculated using Fisher’s exact test. ^✶P<0.05

Table 4: Confusion matrix showing the relationship between background neutrophil infiltration and tumor invasion depth (non-muscle-invasive versus muscle-invasive) in urothelial carcinoma patients.

Depth of invasion	Background neutrophil infiltration			Chi-square
	Low	High	P-value
Muscle invasion (−)	74	22	0.0014✶	10.269
Muscle invasion (+)	22	22

P-value was calculated using Fisher’s exact test. ^✶P<0.05

The utility of scatter plot analysis in metachronous cancer

Finally, we attempted to determine whether scatter plot analysis could be utilized for metachronous cancer. As shown in Figure 8, two of our patients developed metachronous cancer: one 3 years and one 5 years after the initial TUR-Bt. The plot patterns for these patients were all vertical. Interestingly, among the cytologically suspicious cases, Case 73-year 1 and Case 103-year 5 had more tumor cells in the lowest part of the tumor-positive area than the cytologically malignant cases, Case 73-year 3 and Case 103-year 1. These findings suggest that the scatter plot pattern of metachronous tumor cells would be different from that of other tumor cells.

Close

Figure 8:

Scatter plot of the area of cell/cell aggregate and their largest nuclear area in the cell/cell aggregate for metachronous tumor cases. (a and b) Scatter plot of initially developed tumor and secondary developed tumor in Case 73. (c and d) Scatter plot of initially developed tumor and secondarily developed tumor in Case 103.

Export to PPT

DISCUSSION

The previous studies analyzing cytological specimens have primarily focused on nuclear morphology across various samples, including urothelial carcinoma,^[10] pancreatic cancer,^[17] bile duct cancer,^[18] and thyroid cancer.^[11,19] For example, Sakumo et al. found that the nuclear area, cell area, N/C ratio, and nuclear roundness were significantly higher in HGUC compared to reactive urothelial and tubular epithelial cells.^[10] In addition, Poropatich et al. demonstrated that the nuclei of urothelial carcinoma in situ and HGUC were significantly larger than those in reactive urothelial cells.^[20] Momeni-Boroujeni et al. reported significant differences in nuclear area and perimeter between benign and malignant cells in pancreatic cancer.^[17] Collins and Weimholt compared the sizes of individual nuclei and the number of nuclei in bile duct cancer cell clusters, revealing a significant difference in nuclear area between cytologically atypical cases and those suspicious for malignancy.^[18] Chain et al. also noted that the nuclear area was larger in papillary carcinoma and NIFTP compared to benign thyroid cells.^[19] In this study, we focused on the cell/cell cluster area and their maximum nuclear area within the clusters, as we believed that tumor cell clusters exhibited cellular overlapping and higher density with an elevated N/C ratio. In examining cytological specimens, we consider various findings, including cellular overlap and density, as part of the diagnostic criteria in cytology.^[12,21] As cells become malignant, their nuclei enlarge, leading to an increase in nuclear area and cell density. Accordingly, malignant cell clusters are irregularly overlapped compared to benign cell clusters. This finding provides an indicator with which to distinguish between benign and malignant lesions.^[22] However, we did not find any studies that analyzed cytological specimens for overlapping cellular clusters by combining cellular and nuclear areas. In this study, we leveraged the observation that multiple overlapping nuclei in a tumor cell cluster are perceived as a single-nuclear color tone region. This is due to the ambiguous focus of the overlapping tumor cell clusters, which exhibit a high N/C ratio and high cellularity, resulting in the regional connection of nuclear areas. In addition, we could not locate any image analysis studies that applied an algorithm combining the cell/cell cluster area and the maximum nuclear area within those clusters. Therefore, our manuscript is unique and valuable, as our algorithm can distinguish between benign and malignant cell/cell clusters based on different scatter plot patterns. Regarding the use of the Z-stack function in image analysis, while it is commonly employed for analyzing cell images in WSI analysis,^[3,16,23-25] many cytology image analyses do not utilize Z-stacks.^[14] Thus, even though our study did not incorporate the Z-stack function, our research strategy remains relevant and effective.

In this study, we also explored the potential for detecting early-stage cancer by comparing low-grade tumors and high-grade tumors. The scatter plots indicate that many tumor groups (both non-invasive and invasive) did not exhibit the horizontal patterns seen in the non-tumor groups. This implies that urothelial carcinoma can be identified at an early stage in high-grade cases. In addition, our study is significant because our algorithm can detect cancer cells regardless of the depth of tumor invasion. We included only four cases of TURBT specimens with low-grade urothelial carcinoma, focusing on cases classified as “suspicious for malignancy,” “fairly conclusive of malignancy,” and “malignancy.” Future research will need to increase the number of low-grade pTa cases for further study. The scatter plot pattern for low-grade pTa cases is predominantly diagonal, in contrast to the vertical pattern seen in high-grade cases and the horizontal pattern exhibited by non-tumor groups. This indicates not only a difference in patterns between pTa low-grade and pTa high-grade groups but also a difference between pTa low-grade groups and non-tumor groups. Using the scatter plot patterns in this study, it may be possible to predict the presence of tumor components and whether they belong to the low-grade or high-grade groups. The sensitivity for detecting low-grade urothelial carcinoma in urine cytology is low.^[26] Indeed, McCroskey et al. reported that the presence of papillary clusters with fibrovascular core was useful for diagnosing low-grade urothelial carcinoma from normal urothelial clusters, but the appearance of this feature was very low (two out of 98 cases).^[12] In addition, they mentioned that there was no statistically significant difference between the low-grade urothelial carcinoma group and the negative control group for cellular features in multivariate analysis.^[12] In addition, low-grade urothelial carcinoma shows low cellularity and appears as single scattered cells.^[26] The formation of contiguous nuclear regions is influenced by cell density and cell clusters, which explains the fewer vertical patterns observed in the pTa low-grade group.

Finally, we consider the reasons for the significantly lower percentage of tumor-rich areas in the invasive tumor group (pT1-pT4) and the muscle-invasive tumor group (pT2-pT4) compared to the non-invasive and non-muscle-invasive groups. Previous reports have indicated that the number of background neutrophils in urine cytology specimens can serve as an indicator for evaluating the chromatin of tumor cells in urothelial carcinoma.^[27] However, there are no reports that evaluate the number of background neutrophils in urine cytology specifically for urothelial carcinoma. Some studies have noted a correlation between muscle invasion in urothelial carcinoma and the number of neutrophils present in the tumor stroma of histological specimens.^[28] In this study, we propose that the higher number of background neutrophils in urine samples from the invasive tumor group may be due to increased neutrophil infiltration in the stroma, with neutrophils exuding into the urine when tumor cells detach from the tissue. That means, based on our findings, it is possible to anticipate muscle invasion with the findings of increased background neutrophils in the cytological specimens.

SUMMARY

In this study, we introduced two factors not previously utilized in computer-assisted image analysis: cell clusters and the largest connected area of nuclei within those clusters. Our examination of scatter plots revealed that negative cytology cases typically exhibited a horizontal pattern along the cell cluster area axis, while tumor cases showed a vertical pattern along the axis of the largest nuclear area within the cell clusters. This distinction allows for the identification of tumors in urine samples using scatter plots. The algorithm applied in this study aligns with the methods used by cytopathologists and cytotechnologists in their daily practice, focusing on overlapping cell clusters, increased cell density, and a higher N/C ratio in tumor cell clusters. Although we analyzed ambiguous focus tumor cell clusters, we aimed to clarify the characteristics of malignant clusters with high cell density and overlapping clusters compared to normal cell clusters that appear sheet-like and flat. While such ambiguous focus areas are typically overlooked, our study offers valuable insights into how human-utilized algorithm can be integrated into computer-assisted image analysis for future cytological research.