Superpixel Image Segmentation of VISTA Expression in Colorectal Cancer: Implications for Immunotherapy

Colorectal cancer is an overall bad player and accounts for 9% of all cancers. Today, advancements in immune checkpoint inhibition has provided therapeutics for many, but not all cancer patients. This issue is in part due to the tumoral microenvironment; which plays a signicant role in determining response to immune check point therapeutics. This study serves as the rst to evaluate a potent inhibitory checkpoint: V-domain immunoglobulin suppressor of T cell activation (VISTA) and its role in CRC. This was evaluated with both conventional light microscope and superpixel image segmentation. Here we found VISTA expression to be associated with low tumor budding, lower tumor stage, high tumor inltrating lymphocytes, mature stromal differentiation, BRAF mutation status and better disease-free survival in colorectal cancer. When comparing methodologies; superpixel image segmentation better stratied patients into prognostic groups. Anti-VISTA clinical trials are now open and recruiting for patient enrollment for patients with certain advanced solid tumors. Considering raised VISTA expression is associated with improved survival for patients with colorectal cancer: careful, well-designed and robust clinical trials should be pursued in this cancer subtype. cases from these blocks. Hematoxylin and Eosin stained slides were also evaluated for stromal differentiation, tumor budding and tumor-inltrating lymphocytes. Further clinicopathological data was collected from the electronic medical records for each case and analysed in relation to VISTA expression. In exploratory analyses: VISTA expression was compared to multiple variables including: cancer-free survival, age, gender, tumor budding score, pathologic stage, tumor-inltrating lymphocytes (TIL), nodal status, tumor grade, stromal differentiation. Mis-match repair (MMR) and


Introduction
Colorectal cancer is clinically diverse and has a wide spectrum of possible clinical outcomes. It is notable for being one of the most common inherited cancers, although it also occurs sporadically 1,2 . Many patients present with localized disease; while others present with more advanced anatomic extent of disease. Such patients receive more aggressive treatments, often times with immunotherapies; such as those targeting the Program Death-Ligand 1 (PD-L1) axis. For patients with metastatic colorectal cancer with microsatellite instability: Nivolumab has been found to provide clinical response and disease control in the checkmate 142 trial 3 .
However, the world of immunotherapy is not limited to the PD-L1 axis; the tumor microenvironment fosters a complex ecosystem of immune escape. One such marker is v-domain Ig suppressor of T cell activation (VISTA). Unlike PD-L1, VISTA is mainly expressed in immune stromal cells. This includes hematopoietic cells and myeloid cells; as well as naïve CD4 + and Foxp3 + regulatory T cells 4 . NK-like Tcells, myeloid-derived suppressor cells (MDSCs), such as: macrophages, granulocytes, dendritic cells and immature myeloid cells: work to regulate the immune microenvironment in cancer 5 .
VISTA is on the cutting edge and as of now; studies have described the suppressive effect of VISTA, with a presumed e cacy for anti-VISTA therapy. However, VISTA is highly controversial: it acts as a ligand on antigen-presenting cells, while also serving as a receptor on T cells 6 . Clinical trials with anti-VISTA are ongoing (Clinicaltrials.gov) and studies correlated with the expression and VISTA as a poor prognostic signature will be important in justifying its use. Apart from obvious differences in survival, evaluating VISTA in the setting of the tumoral microenvironment will also be interesting. Immune markers are often associated with tumor in ltrating lymphocytes and molecular alteration such as microsatellite tumoral instability.
However, other correlations also exist: with the role of tumor budding and immature stromal differentiation yet to be discovered. In breast cancer, immature stromal differentiation: de ned by the presence of myxoid stroma, was found to be inversely associated with PD-L1 expression 7 . In colorectal cancer, stromal differentiation has already been validated as signi cant prognostic signature by numerous studies [8][9][10][11] .
Recent advancements have been made in image analysis and segmentation. Including the use of Qupath as an open-source solution for computational pathology and whole slide image analysis 12 . Qupath offers many applications for image analysis, including superpixel segmentation; which has become a popular approach in computer vision: often used to reduce image complexity for digital analysis 13 . Here we will utilize superpixel image segmentation for the purpose of evaluating VISTA expression.
This study is the rst to evaluate VISTA expression in colorectal cancer; therefore, its relationship to the tumoral microenvironment remains largely speculative.

Study Design:
A total of 226 cases of colorectal carcinoma diagnosed in our health system were retrospectively analyzed and cases were collected from 2012-2017. Cases were only selected if they were primary resection specimens, had nil adjuvant therapy and availability of su cient non-frozen, formalin-xed, para n-embedded tumor material for research purposes. One representative block was selected per case from a single-slide containing the largest portion of tumor. VISTA immunohistochemical (IHC) expression was evaluated on 226 cases from these blocks. Hematoxylin and Eosin stained slides were also evaluated for stromal differentiation, tumor budding and tumor-in ltrating lymphocytes. Further clinicopathological data was collected from the electronic medical records for each case and analysed in relation to VISTA expression. In exploratory analyses: VISTA expression was compared to multiple variables including: cancer-free survival, age, gender, tumor budding score, pathologic stage, tumorin ltrating lymphocytes (TIL), nodal status, tumor grade, stromal differentiation. Mis-match repair (MMR) status, Ki-ras2 Kirsten rat sarcoma viral oncogene homolog (KRAS) and B-Raf Proto-Oncogene (BRAF).

Immunohistochemistry:
All the staining was performed on Formalin-xed and para n-embedded tissue blocks. All staining was

Superpixel Analysis of VISTA:
QuPath is an open source for whole slides image analysis 12 . Studies have shown a high accuracy in QuPath image analysis as descried in Bankhead et al. In clinical practice, an individual pathologist manually interprets staining expression; however, this is prone to observer variation. In our study, we adopted QuPath and chose to utilize superpixel image segmentation to evaluate VISTA expression in addition to manual analysis. IHC slides were scanned using Aperio and QuPath version 0.2.1 Superpixel method is used to quantify the staining. 10x hotspots were identi ed from the whole slide image by a surgical pathologist. Superpixel automatically groups pixel similarity between different cellular populations as described in Wang et al 13 . Pixel similarity was determined by their red, blue and green values 13 .
Hotspots were used to train the Superpixel method. For this, areas of interest were selected and categorized as "Tumor", "Stroma" and "VISTA". Heatmap was generated after training and the percentage is calculated spontaneous. Once the Superpixel has been optimally trained, all hotspots were calculated under the same Superpixel image segmentation model. We selected Superpixel method over automatic cell quanti cation for following reasons rst described in Wang et al and also been observed in our study. The RBG values for our hotspots remained stable and once we trained the system, it was able to pick up different RBG values from new cases. Overall, it allowed for a faster and stable quanti cation, with minimal changes in behavior. Traditional cell quanti cation recognizes cellular expression; however, superpixel segmentation makes calculations based on RBG values, resulting in more subcellar information.

Stromal Differentiation:
For stromal differentiation, scoring was based on the 3-tier grading system proposed by Ueno et al 9 .
( Fig. 1). We analyzed the extramural desmoplastic front at low magni cation. We categorized keloid-like collagen as thick hyalinized bundles comprised of hypocellular eosinophilic hyalinized collagen, similar to that of a keloid scar. Myxoid stroma was de ned as an amorphous stromal substance made of amphophilic material with a basophilic to grey extracellular matrix. This was usually intermingled with randomly oriented hyalinized collagen. As in Ueno et al stroma grading system, stroma was regarded as immature when brotic stroma with myxoid changes was observed. In a tumor with no myxoid stroma, stromal differentiation was categorized as intermediate when keloid-like collagen was intermingled with the mature stroma. Finally, stroma differentiation was regarded as mature when the brotic stroma did not contain myxoid stroma or keloid-like collagen and most commonly comprised of ne mature collagen bers strati ed into multiple layers.
Tumor Budding: For tumor budding assessment, a detailed search was done for the area having the highest grade of tumour budding. After that, the counting of the buds took place in the hotspot region (20 × objective lens). According to ITBCC protocol, the tumor budding is graded using 3-tier grading system as follows: Bd1: 0-4 buds, Bd2: 5-9 buds and Bd3: 10 or more buds.

VISTA Expression:
The overall expression of VISTA on manual analysis was 22%, while 21 cases showed 0% expression and 6 cases showing more than 90% expression. The overall expression of VISTA on superpixel was 20%, while 15 cases showed 0% expression and 3 cases showing more than 90% expression. Pearson's correlation coe cient demonstrated good correlation between manual and superpixel analysis (r = 0.92), shown in Fig. 2. Heatmaps for VISTA expression between manual and superpixel analysis can be seen in Fig. 3.

VISTA Expression and disease-free survival:
All of the 226 patients are followed with a mean follow up time of 1054 days. Nonlinear regression found the optimal cutoff for VISTA staining and survival to hover at the 20% expression mark for both manual and superpixel analysis. After setting up the cutoff of high positive value for 20%, we found the correlates of high VISTA expression with favorable survival rates (P < 0.05), as shown in Fig. 4.
For manual analysis: the mean disease-free period for VISTA expression > 20% is 789 days, 197 days longer than those with VISTA expression < 20%. During the follow-up period, 85% cases with VISTA expression > 20% were disease free compare to the 70% disease free rate for those with VISTA expression < 20%. For superpixel, the mean disease-free period for VISTA expression > 20% is 787 days, 164 days longer than those with VISTA expression < 20%. During the follow-up period, 90% cases with VISTA expression > 20% were disease free compare to the 65% disease free rate for those with VISTA expression < 20%.
VISTA Expression and the Clinicopathologic Pro le: For tumor budding, high tumor budding was found to be associated with low VISTA expression (P = 0.01) on superpixel analysis. High pathologic tumoral stage was also found associated with low VISTA expression on superpixel analysis (P = 0.04). High tumor-In ltrating lymphocyte scoring was found to correlate with higher VISTA expression on superpixel analysis (P = 0.01). When analyzing stroma differentiation, low VISTA expression was found to correlate with immature stroma differentiation by superpixel analysis (P = 0.02). Mature stroma had the highest expression, with a mean positivity of 24%, whereas immature stroma had a mean positivity of 17%. Age, gender, lymph node status and tumor grade did not corelate signi cantly with VISTA expression (P > 0.05). Results from Fisher exact analysis and T test can be seen in Table 1 and Fig. 5.  Table 2 and Fig. 6.

Discussion
Overall, immune checkpoint targets have been studied extensively in colorectal cancer. However, VISTA is a relatively novel immune checkpoint and its signi cance was previously investigated in both esophageal and gastric adenocarcinomas 15,16 . For VISTA and colorectal cancer, our ndings also suggest VISTA to be a clinically signi cant biomarker.
Firstly, we found VISTA correlates with stromal differentiation, a novel concept described recently that associated with prognosis outcomes in colonic adenocarcinoma 8 . We found higher VISTA expression to correlate with more mature stroma: well depicted by superpixel image segmentation (Fig. 7).
Tumor budding and TILs were also found to be associated with VISTA expression, suggesting that a histologic phenotype exists which support VISTA expression in colorectal cancer. Our study suggests this phenotype includes: high VISTA expression, low tumor budding, mature stromal differentiation, lower tumor stage and longer survival.
Regarding the prognostic ndings in our study, it is important to understand that VISTA is a complicated immune check point regulator. One which is present on both CD4 + and Foxp3 + regulatory T cells as well as myeloid-derived suppressor cells MDSCs. In esophageal adenocarcinoma, signi cant VISTA expression was found on CD4 + T cells: which resulted in improved prognostic outcomes 15 . Such prognostic ndings were also found in our study.
As a hypothesis, it is possible that survival outcomes in our studies are due to VISTA and its role in converting naïve T cells into FoxP3 expressing T cells, which has been demonstrated before 17 .
Importantly, FoxP3 has been shown to improve survival in patients with colorectal cancer 2, 18 .
FoxP3 is known to be a transcriptional repressor of the proto-oncogene SKP2, which regulates the cell cycle in the G2/M phase 2,19 . Inhibited FoxP3 expression results in the overexpression of SKP2 and cell proliferation.
Saying this, it is possible that VISTA in this setting may be immune-protective for patients with colorectal cancer. On the other hand, VISTA expression in melanoma was found to be a poor prognostic marker 20 .
In essence, this may represent the multifaceted role of VISTA for different tumor subtypes. Clinical trials will have the ultimate say in determining therapeutic e cacy; however, anti-VISTA therapies may not be fruitful in this cancer subtype.
In our study, we utilized Qupath: an open access image analysis software. While doing so, we compared the results with that of standard manual analysis. Overall, superpixel analysis outperformed its manual counterpart. The separation of prognostic cohorts was more statistically signi cant, and superpixel correlated better with the overall clinicopathologic pro le.
In computer vision the use of image segmentation dissociated image segments into colored sets of pixels (super-pixels). This allows for more rudimentary image representations, which allows for ease of analysis. As this sounds ideal for the purpose of biomarker expression in the eld of pathology, we decided to super-pixelate VISTA hotspots for the purpose of classi cation.
There has been some talk of superpixel analysis in the eld of medicine. In radiology, brain and breast radiography has been targeted as a novel segmentation method with improvements over current state of the art 21,22 . In the eld of histopathology, it has been shown to improve segmentation results for deep learning based topologies in cervical intraepithelial neoplasia 23 . In future, superpixel image segmentation could be integrated into the ever-evolving digital pathology ecosystem and becomes ubiquitous to the analysis of biomarker expression. This will facilitate the evolution of the tradition to fully digital pathology work ow.
Regarding the use of hotspots for image analysis in our study; studies have found that hotspot analysis to outperform global scoring. Robertsona et al. found that hotspot scoring had double the potential for predicting recurrence free survival, in the setting of Ki-67 immunohistochemistry and breast cancer 24 .
The analysis of hotspots may better correlate with immune activity and future applications in deep learning will allow for automatic detection of hotspots from whole slide images: alleviating the need for manual evaluation 25 .
There are several pitfalls to our study and it is important to consider that this was a retrospective study with the potential for bias. Secondly, we did not perform molecular testing for VISTA expression, which would have been more thorough. Thirdly, we did not perform multiplex immunolabeling in this study, which would have allowed us to identify VISTA expression in different immune cell subtypes. This would have been particularly useful for identifying CD4 + and FoxP3 expressing T cells.
One novel nding is the association for BRAF mutations and higher VISTA expression. This has never been reported; however, a recent publication has shed light onto these phenomena, where BRAF inhibition in BRAF mutant patients increased FOXP3, while downregulating VISTA expression 26 . Our ndings support that the BRAF mutant mutation to be associated with higher VISTA expression. While the role of BRAF in the tumoral microenvironment is not well known; it is possible that BRAF may have therapeutic effect through modulation of VISTA expression. The relationship between BRAF and VISTA should be explored in future studies. Today, BRAF inhibitor therapy happens to be ubiquitous to chemotherapy, particularly for patients with melanoma 27 .
For VISTA, there are many questions still remain. Future studies examining the therapeutic e cacy of anti-VISTA, in combination with other immune check point inhibitors, will also be useful.
Looking forward, VISTA could be on the cutting edge of an immunotherapy revolution. However, targeting VISTA in an inhibitory fashion should be performed with caution. It may be immune-protective for patients with colorectal cancer.     Decreased VISTA Expression and Immature Stromal Differentiation. VISTA, V-domain immunoglobulin suppressor of T cell activation.