Prognostic significance of tumor budding in pancreatic carcinoma: Digitalized image approach evaluation using artificial intelligence.

Introduction

Pancreatic carcinoma (PC) is an aggressive malignancy, representing the fourth leading cause of cancer-related deaths in developed countries, with an increasing incidence projected to make it the second leading cause by 2030.¹ It is often diagnosed at advanced stages (stages III-IV in over 80% of cases, where only 15-20% of patients are candidates for curative surgery.² Despite complete resection as the primary treatment for resectable disease, PC exhibits rapid progression and high recurrence rates (up to 90% within 5 years), even in node-negative cases, resulting in a 5-year survival rate of approximately 13% overall and 44% for localized, resected tumors. Although tumor resectability and stage remain the most relevant prognostic factors, there is an increasing need to focus on novel histo-prognostic factors that would enable better risk stratification for patients.¹

Tumor budding (TB), defined as the presence of isolated single cancer cells or clusters of up to four cancer cells at the invasive tumor front, has been recognized as an emerging marker of aggressiveness related to the epithelial to mesenchymal process in patients with colorectal cancer. Thus, TB has been introduced as a routine prognostic marker in colorectal cancer to stratify patients for adjuvant chemotherapy for stage II tumors.^3,4

In PC, although it has been reported that TB has a clear association with adverse prognosis, TB is still not systematically reported by pathologists, and there are no clearly defined recommendations on TB counting in PC.^5–8

This study aimed to assess the tumor budding score in PC using artificial intelligence and to explore the association of TB clinicopathological prognostic factors.

Methods

Tumor budding counting

Pathological HE slides analysis

All hematoxylin and eosin (HE)-stained slides were reviewed by two pathologists to select the most representative slides for tumor budding assessment.

Tumor budding was assessed according to the recommendations of the International Tumor Budding Consensus Conference (ITBCC) for colorectal cancer.⁹ TB was identified as a single tumor cell or a cluster of <5 cells. For each selected HE-stained slide, hotspot areas either at the invasion front or within the tumor center were identified at 10-fold magnification. TB was counted by two pathologists over a field of 0.785 mm² at 20-fold magnification.

Digital assessed approach

The slides that were selected for TB counting at 20-fold magnification were digitalized using NIS image scope software, which was connected to a Nikon microscope (Eclipse Ni-U) ( Figure 1).

Figure 1. Screenshot of digitalizing an HE X 20 slide on NIS software.

These images were exported in GIF format and then uploaded to open-source software QUPATH (version v0.4.3, https://qupath.github.io/) for digital pathology.¹⁰ TB was assessed and evaluated using a semi-automated method according to the methodology of Budeau et al.¹¹ In this study, the authors assessed TB in a cohort of 92 patients with intrahepatic cholangiocarcinoma. Firstly, Tumor budding was identified in one tissue slide on the basis of the recommendations of the International Tumor Budding Consensus Conference 2016.⁹ The HE slides were then digitalized and all images were analyzed in QuPath 16 (Version 0.1.2) (https://qupath.github.io/). Ten rectangles of H&E stained tissue slides were evaluated for both, the tumor-host interface and the tumor center. Each rectangle was standardized for an area of 0.785 mm² as recommended. The function “cell detection” facilitates the manual differentiation of tumor buds from larger tumor groups by detecting individual cells and only having to assess the number of cells.¹¹ In the present study, according to this methodological approach, we used the annotation functionality of QUPATH to highlight tumor cells (red color). Each 20-fold magnification digitalized slide was segmented into five rectangles, each rectangle corresponded to an area of 0.785 mm.^4,10 We used the “cell detection” functionality to circumscribe tumor cells based on nucleus sigma between 3 and 8. The TB score was counted manually for each rectangle, and the final TB score for each case was defined as the average of the five rectangles ( Figure 2a, b and c).

Figure 2. (a, b and c) Detection of single tumor cells and small clusters on QUPATH software.

Classification of tumor budding

For both morphological and semi-automated methods, the final TB score for each case was categorized into four groups according to the recommendations of ITBCC⁴:

• ✓ BD0: no budds

• ✓ BD1: 1-4 Budds

• ✓ BD2: 5-10 Budds

• ✓ BD3:>10Budds

→ The final TB score is calculated as follows:

• ➢ Low: BD0-BD1

• ➢ High: BD2-BD3

Results

25 cases patients were included in this study. The mean age of the patients was 62.3years old the male-to-female ratio was 2.57. Alcohol consumption and smoking were found in 36% and 56% of the patients, respectively. Clinical signs included abdominal pain (96%), loss of body weight and condition (80%), jaundice (72%), transit disorder (28%), dark urine (64%), discolored stool (44%), pruritus (36%), and fever (16%). On physical examination, a palpable gallbladder was found in 36% and ascites in 12% of patients. Laboratory tests revealed anemia in 64%, elevated CA19.9 in 76%, cytolysis in 52%, and cholestasis in 76%. Surgery was performed in 80% of the cases: cephalic duodeno-pancreatectomy (75%), caudal pancreatectomy (20%), and double bypass (5%). Neoadjuvant therapy was performed in 12% of patients, adjuvant treatment in 56%, and palliative care in 28%. Local recurrence occurred in 44% of patients, distant metastases in 38.9%, and death in 72%. The pathological findings are summarized in Table 1.

Tumor budding was found in 100% of cases using the morphological method and 84% using the digitalized approach. The number varied from 1 to 37 (mean: 8.04, median:4) using the morphological method and from 0 to 19 (mean: 5.92, median: 6) using the QUPATH software. Tumor budding counts are summarized in Table 2.

A high TB score (BUDD2-3) was found in 48% of the cases using the morphological method and 56% using the semi-automated QUPATH method. No statistically significant differences were observed between the two methods (p=0.589).

Multivariate analysis

On multivariate analysis, tumor grade, vascular invasion, and tumor budding affected overall survival (p=0.04, p=0.07, p=0.016, respectively) ( Figures 3, 4, 5).

Figure 3. Tumoral budding impact on overall survival (time of death).

Figure 4. Vascular invasion impact on local recurrence.

Figure 5. Tumor grade impact on recurrence.

Discussion

In the present study, tumor budding was found in 100%of cases using morphological methods and 84% of cases using the digitalized approach. The difference observed between both methods suggests that using a semi-automated approach may offer greater analytical precision by reducing false positives. These cases likely corresponded to atypical mesenchymal fibroblastic or myofibroblastic cells or poorly differentiated tumor glands that were mistaken for isolated tumor cells on microscopic evaluation.

Overall, according to our results, tumor budding is observed at a high frequency in pancreatic carcinoma, which is consistent with previous reports indicating that TB is reported in approximately 85-100% of specimens with pancreatic carcinoma.^5,7,8 This fact raises the hypothesis that TB is a relatively constant finding that could also be a pervasive feature of pancreatic carcinoma, especially in biopsy specimens.⁵

In our study, a high TB score was found in 48% morphologically, and in 56% using QUPATH. Hence, using the QUPATH software reduced the number of cases classified as high BUDD3 with the morphological approach. This is due to the software’s precision in distinguishing between isolated tumor cells or clusters of fewer than 5 cells, which meet the definition of tumor budding, as opposed to poorly differentiated carcinoma clusters or remnants of neuroendocrine islets that were misdiagnosed as TB on HE slides examination and were probably false positives. Indeed, the annotation option in QUPATH software allows for the identification and deselection of false positives.

In the present study, the rates of high TB were slightly lower than that reported in previous studies reporting a high TB score in about 56 to 80% of cases.^5,7,8 These disparities may be partly explained by the differences in TB counting methods, the surface of the HPF, and the use of immunohistochemistry in some studies to identify CK+-stained tumor cells. TB is defined as single cells or clusters of <5 cells at the tumor invasion front.^4,9 This quantifiable histological feature has gained attention over the last 10years and has proven its prognostic value in many cancers. The International Tumor Budding Consensus Conference (ITBCC) proposed a scoring system for tumor buds in colorectal carcinomas that was later applied to other cancers such as hepatocellular, oral squamous cell, and bladder cancers.^9,11 However, there is still no standardized method for reporting tumor budding in PC, and various approaches are available to count tumor buds either on HE slides or using pathological image analysis softwares.^11,12 In the present study, we first assessed tumor budding in pancreatic carcinoma specimens according to the recommendations of the ITBCC at 20-fold magnification for a field of 0.785 mm², either at the tumor front or within the tumor center. Then, similar to the published study of Budeau et al.,¹¹ we extrapolated the ITBCC counting approach on an optic microscope to a digital image analysis system using QUPATH, an open-access software for digital image analysis. QUPATH is completely free, which enables pathologists in low-income countries to have access to digital pathology analysis.¹⁰ The tumor cell detection functionality of the QUPATH software facilitates distinguishing between tumor cells and small clusters from a larger group of cells. In addition, compared to “analog” microscopy, QUAPTH offers the possibility of re-evaluating tumor buds at anytimes higher than those of other pathologists.¹¹ Although a high TB score was found in 48% of patients using the first method compared with 56% using the digitalized approach, the difference was not statistically significant (p=0.589). Hence, QUPATH could be an interesting, costless, and accurate alternative for pathologists, which would considerably reduce the work time and facilitate TB counting. This approach is particularly relevant for enhancing reproducibility and efficiency in TB scoring, potentially standardizing its use in clinical pathology worldwide, especially in low-resource settings.

According to the recommendations of ITBCC, the TB score is categorized as a three-tier grading system for patient risk stratification.^9,11 In the present study, similar to the study published by Tanaka et al.,¹² we first graded the TB score as low (BUDD1), intermediate (BUDD2), and high (BUDD3) but later combined BUDD2 and BUDD3 groups into high TB scores for statistical analysis purpose. Consequently, we compared the high and low TB score groups using clinicopathological and survival data.

Using the morphological method results, we demonstrated a statistically significant (or nearly significant) association between high tumor budding and advanced age >72 years (p=0.03), ductal subtype (p=0.07), and advanced-stage pT>pT2 (p=0.08). These results are consistent with those of previous studies that demonstrated an association between high TB and prognostic factors such as poor differantiation, lympho-vascular invasion, peri-neural invasion. Regarding prognosis, high TB was independently associated with worse OS (HR 2.35), reflecting its role in EMT, promoting metastasis and recurrence.¹³ Multiple meta-analyses and systematic reviews have consistently established high-grade TB as a robust, independent predictor of poor outcomes in PDAC, including reduced overall survival (pooled HR 2.13, 95% CI 1.63-2.78) and disease-free survival (pooled HR 1.85, 95% CI 1.45-2.36), even after adjusting for stage, grade, and lymph node status.^14–16

In some studies, high TB was linked to adverse clinicopathologic features such as lymphovascular invasion, perineural invasion, higher tumor grade, and nodal metastasis, which contribute to its prognostic impact.^13,17 Recent studies, including those in neoadjuvant therapy settings, showed that TB remains prognostic, with high budding indicating increased risk of recurrence and metastasis, supporting its use for identifying patients who may benefit from intensified adjuvant therapies or closer surveillance.^5,18–20 Furthermore, TB’s association with EMT markers suggests it represents a biologically aggressive phenotype, advocating for its routine inclusion in pathology reports to enhance risk stratification and personalize treatment in PDAC.^14,21 In a study published by O’Connor et al.,⁷ high-grade budding (> 10 buds in 10 HPFs) was associated with a high tumor grade, lymphovascular invasion, and perineural invasion. In our study, 53.8% of patients with lymph node invasion or advanced stage had a high TB score; likewise, 76.9% of patients with perineural invasion had a high TB score. However, these findings were not statistically significant (p=0.53, p=0.32, and p=0.53, respectively), which could be related to the small sample size in our study.

Of note, regardless of the method used, many other studies on pancreatic cancer did not reveal a statistically significant association between TB and prognostic factors.^5,7,22–24

In the present study, the authors described a statistically significant association between tumor budding and overall survival (p=0.016) in the multivariate analysis after adjusting for other significant variables. This finding is consistent with previous reports that support the prognostic value of the TB score in pancreatic carcinoma for better patient stratification and treatment guiding.^5,7,8 Although these results are promising, there are still many discrepancies in the TB counting methods. First, some authors only considered TB at the tumor-host interface⁷; however, there is increasing evidence that intra-tumor and peri-tumor buds have comparable prognostic significance.^{9,11–13,18,22–26} Second, some authors suggested that tumor budding evaluation is more accurate and reproducible using immunohistochemistry with anti-CK antibody³; however, in a multicenter study, Hacking S et al,²⁷ demonstrated that although immunohistochemical staining facilitates the detection of tumor cells, it has comparable intra- and inter-observatory reproducibility to the HE slides TB counting approach. Finally, in the era of artificial intelligence, using digital pathology for TB assessment is an interesting alternative approach that could considerably reduce the examination time and offer better accuracy and reproducibility. In this context, many studies have demonstrated a high range of diagnostic concordance (90-99%) between digital slides and conventional glass slides.^28–30 Many digital pathology image software packages have been developed; however, access to these platforms remains difficult for some pathologists.^31,32 Hence, QUPATH software could be an interesting alternative for integrating digital pathology in routine practice. A consensus on tumor budding counting and reporting on digitalized slides in pancreatic carcinoma is necessary to definitively integrate TB in pathology reports.

Conclusions

Our results demonstrated that tumor budding either assessed manually on HE slides or semi-automated on digitalized images is a relatively constant finding in pancreatic carcinoma, with a high score in about 50% of cases. Our findings also showed a high concordance of both methods in TB assessment, supporting the benefits of integrating digital pathology in routine practice. Finally, our results provide further evidence for the potential prognostic value of TB in pancreatic carcinoma. However, our study has some limitations, such as its retrospective design and small sample size. Further studies on TB counting using QUAPATH in pancreatic carcinoma are needed to integrate this method into routine practice for better risk stratification.

Ethics approval

This research was conducted following the ethical guidelines outlined by the Ethics committee of the Internal Security Forces Hospital (Obtained on 1 ^st December 2023, approval number 12/23). All procedures involving human tissues were approved by the committee and were performed in accordance with the ethical standards laid down in the 1964 Declaration of Helsinki and its later amendments as well as the National Medical Code of Ethics (Title VI, Article 99 to 111). Verbal informed consent was obtained at admission from all individual participants included in the study. The majority of patients were illiterate, unable to read or write, therefore verbal consent was preferred. Confidentiality and anonymity of participants were strictly maintained throughout the study. Any potential conflicts of interest have been disclosed and managed appropriately. Confidentiality and anonymity of participants were strictly maintained throughout the study. Any potential conflicts of interest have been disclosed and managed appropriately.

Link to National Medical Code of Ethics: chrome-extension://efaidnbmnnnibpcajpcglclefindmkaj/ http://www.atds.org.tn/Decretdeontologiemedicale93.pdf

Authors contribution

Sarra Ben Rejeb: Conceptualization Investigation; Methodology Writing Validation;

Jasser yaacoubi: Data curation; Formal analysis;