Task Call: Creation of Large-Scale Annotations for Tumor Budding and Poorly Differentiated Cluster in Colorectal Cancer

TL;DR:

Tumor budding (TB) and poorly differentiated clusters (PDCs) are crucial prognostic factors in colorectal cancer (CRC). This project aims to create a large, diverse, and publicly available dataset of annotated histopathology images to develop AI algorithms for detecting and quantifying TB and PDCs. The article outlines the clinical importance of TB and PDCs, current challenges in manual assessment, project goals, related work in computational pathology, and detailed guidelines for dataset preparation and annotation. Contributors will be rewarded with ownership stakes in the annotation dataset, and this dataset will be made freely available for non-commercial use on HuggingFace, along with an AI inference model and an online service.

Background and Clinical Importance

Colorectal cancer (CRC) remains a significant global health challenge, with an estimated 1.9 million new cases and 935,000 deaths in 2020 [1]. While traditional staging methods provide valuable prognostic information, there is a growing recognition of the importance of additional histological features in refining risk stratification and treatment planning. Two such features, tumor budding (TB) and poorly differentiated clusters (PDCs), have emerged as crucial prognostic factors in CRC.

Tumor budding is defined as the presence of single tumor cells or small clusters of up to four tumor cells at the invasive front of carcinomas [2]. Poorly differentiated clusters are defined as clusters of ≥5 cancer cells that lack glandular formation [3]. Both TB and PDCs represent a continuum of the same biological process — the dissociation of cancer cells at the invasive front [4].

Figure 1: Example of tumor budding (a) defined as single tumor cells or tumor cell clusters at up to four cells. Example of poorly differentiated clusters (b) defined as five tumor cells or more. DOI: 10.1038/modpathol.2017.46

The clinical importance of TB and PDCs in CRC management cannot be overstated:

1. Prognostic Value: Both TB and PDCs are independent prognostic factors in CRC, providing valuable information beyond traditional TNM staging. High-grade TB and high-grade PDCs are associated with poor prognosis, including increased risk of lymph node metastasis, distant metastasis, and reduced overall survival [5, 6].
2. Risk Stratification: TB and PDC assessment helps stratify patients into risk categories, particularly in stage I and II CRC. This stratification is crucial for identifying high-risk patients who might benefit from more aggressive treatment or closer follow-up, even in early-stage disease [7].
3. Treatment Decision-Making: In pT1 CRC, the presence of high-grade TB can influence the decision between endoscopic resection and surgical intervention. For stage II CRC patients, who traditionally do not receive adjuvant chemotherapy, the presence of high-grade TB or PDCs may indicate a need for additional treatment [8, 9].
4. Personalized Medicine: TB and PDC assessment contribute to a more personalized approach to CRC management, allowing treatment intensity to be tailored to individual patient risk [10].
5. Surveillance Planning: Patients with high-grade TB or PDCs may require more intensive post-treatment surveillance due to an increased risk of recurrence [11].

Figure 2: Kaplan-Meier estimates of the relapse-free survival (RFS) rate in patients with colon cancer according to grade of tumor budding. 10.1200/JCO.18.02059

Despite their clinical importance, the current manual assessment of TB and PDCs faces several challenges including [12–15] :

1. Time-consuming nature: Manual TB and PDC assessment requires careful examination of large tissue areas at high magnification, which is labor-intensive and limits the number of cases that can be evaluated.
2. Inter-observer variability: Despite established guidelines, there remains substantial variation in TB and PDC assessment between pathologists, potentially leading to inconsistent prognostication and treatment decisions.
3. Need for specific expertise: Accurate identification of TB and PDCs requires specialized training and experience, which may not be universally available, especially in resource-limited settings.

Project Goals and Significance

Given the clinical importance of TB and PDCs and the challenges in their assessment, there is an urgent need for automated, AI-assisted methods for their detection and quantification. This project aims to create a large, diverse, and well-annotated dataset of histopathology images for the development and validation of AI algorithms to assist in detecting and quantifying of both tumor budding and PDCs in CRC.

The creation of this dataset is of utmost importance for several reasons:

1. Improved Prognostication: AI-assisted TB and PDC assessment can significantly improve CRC prognostication, helping clinicians make more informed decisions about patient management and follow-up.
2. Standardization: A comprehensive dataset will contribute to standardizingTB and PDC assessment, reducing inter-observer variability and improving consistency in CRC staging across institutions.
3. Workflow Efficiency: Automated detection and quantification can reduce the time required for assessment, facilitating integration into routine pathology workflows and potentially increasing the utilization of these important prognostic factors.
4. Accessibility: AI-assisted assessment could make TB and PDC evaluation more widely available, even in settings with limited pathology expertise.
5. Research Advancement: A well-curated dataset will enable further research into the biological significance of tumor budding and PDCs, and their relationship with other prognostic factors in CRC.

A well-curated, large-scale dataset can not only provide a foundation for developing AI models that can perform rapid, automated TB and PDC assessments but also serve as a benchmark for assessing and improving consistency in TB and PDC evaluation.

Related Work in Computational Pathology

While computational pathology has made significant strides in various aspects of cancer histopathology analysis, work specifically focused on the automated detection and quantification of tumor budding (TB) and poorly differentiated clusters (PDCs) in colorectal cancer remains limited.

Current TB and PDC detection approaches primarily utilize Convolutional Neural Networks (CNNs). Several studies have explored the use of CNNs for TB detection, with varying architectures and results. Banaeeyan et al. used a SegNet-based model, achieving a mean intersection over union (IoU) of 0.49 [16]. Bergler et al. employed an AlexNet-based approach, reporting high sensitivity (0.934) but lower precision (0.068) [17]. Bokhorst et al. achieved a maximum F1-score of 0.36 with a VGG16-based model [18]. More recently, Lu et al. reported high accuracy (0.89) and sensitivity (0.94) using a Faster R-CNN architecture [19].

Some studies have used a combination of classical image processing techniques and deep learning. Weis et al. used color deconvolution, k-means clustering, and a CNN for TB evaluation [20]. Fisher et al. developed a semi-automated method using QuPath software for TB assessment [21].

Regarding PDC detection, Pai et al. developed a deep learning algorithm to quantify both TB and PDCs, reporting stronger correlations with lymph node metastasis for combined TB/PDC measures compared to TB alone [22].

Despite these advancements, current work in this field faces several limitations:

1. Limited proprietary datasets: Most studies use small, single-institution datasets, which limit the generalizability of the models. Those datasets are not made publicly available for the pathology community.
2. Lack of standardization and annotation QC: There is significant variability in the definition and annotation of TB and PDCs across studies.
3. Focus on single features: Many models focus solely on TB detection without considering PDCs or other related histological features.
4. Validation gaps: There is a lack of large-scale, multi-institutional validation studies to demonstrate clinical utility.

Given these limitations, there is a clear need for a more comprehensive, holistic approach to TB and PDC assessment in colorectal cancer. A new model should simultaneously detect and quantify both TB and PDCs, recognize their biological continuum, integrate assessment of related histological features, utilize larger multi-institutional datasets with cross-validated annotations, provide interpretable results, and offer flexibility to work with various staining/scanning methods.

Dataset Preparations

We will carefully select cases from public digital pathology archives to create a high-quality dataset for AI model development in tumor budding (TB) and poorly differentiated cluster (PDC) detection and quantification. This approach ensures a diverse and representative dataset while leveraging existing resources. The selection and preparation process will follow these guidelines:

1. Data Sources: Identify and utilize reputable public digital pathology archives, namely: a. TCGA-COAD | The Cancer Genome Atlas b. CPTAC-COAD | The Clinical Proteomic Tumor Analysis Consortium Colon Adenocarcinoma Collection
2. Case Selection Criteria:

• Invasive Front Visibility:
  -Select cases where the invasive front of the tumor is clearly visible and well-represented in the whole slide images (WSIs).
  -Ensure the invasive front is sufficiently extensive for meaningful TB and PDC analysis.
• Diversity of CRC Cases:
  -Include a range of colorectal cancer grades: well-differentiated, moderately differentiated, and poorly differentiated.
  -Represent various stages, with a focus on early-stage diseases (Stage I and II).
• Tumor Budding and PDC Levels:
  -Ensure representation of cases with different levels of tumor budding:
  1. Low (0–4 buds per hotspot/0.785 mm(2)corresponding one 20x field)
  2. Intermediate (5–9 buds per hotspot/0.785 mm(2)corresponding one 20x field)
  3. High (≥10 buds per hotspot/0.785 mm(2)corresponding one 20x field)
  -Include cases with varying PDC grades:
  1. Grade 1 (<5 PDCs)
  2. Grade 2 (5–9 PDCs)
  3. Grade 3 (≥10 PDCs)
• Focus on Early-Stage Diseases:
  -Prioritize cases of Stage I and II colorectal cancer.
  -Include a smaller proportion of Stage III cases for comparison and model robustness.
  -Limit the inclusion of Stage IV cases, as TB and PDC assessment is less relevant in metastatic disease.

Figure 3: Examples of different tumor budding grades (hotspot, 0.785 mm2) at the invasive front of colorectal cancer based on the ITBCC 2016. (a): Bd 1 (low), (b): Bd 2 (intermediate) and ©: Bd 3 (high). DOI: 10.1038/modpathol.2017.46

Annotation Platform Registration

Stage 1: Annotation by Two Pathologists

a. Go to https://healthcare.codatta.io/. Use your email to log in the annotation system, remember to check your spam box if you cannot find the verification code.

b. Once logged into the system, please edit your personal information as accurately as possible. Note that only senior level pathologists can validate the annotations.

c. Then, click on “Annotation” on the left menu, and choose “colon” to get in.

d. For “Unannotated” and “In progress” cases, click on “Annotation” to annotate.

• Annotate the representative invasive fronts using the Polygon tool.
• Annotate a representative hotspot with the highest density of tumor buds and PDCs using the specific circle tool (0.785 mm2).

• Annotate all tumor buds (1–4 cells) using the circle tool.
• Annotate all PDCs (≥5 cells) using the circle tool.

• Count and record the total number of TBs and PDCs in the hotspot, and do grading on TB and PDC. Record the numbers in educational notes of the hotspot.
  -Assign TB grade:
  a. Bd1 (Low): 0–4 buds
  b. Bd2 (Intermediate): 5–9 buds
  c. Bd3 (High): ≥10 buds
  -Assign PDC grade:
  a. Grade 1: <5 PDCs
  b. Grade 2: 5–9 PDCs
  c. Grade 3: ≥10 PDCs

Stage 2: Validation and Final Adjustment (Performed by experienced pathologists)

1. Review of Annotations:

• Click on the name of each annotator to see their work
• Examine hotspot and all annotations of each annotator and give scores.
• Verify the accuracy of TB and PDC identification.

2. Grading Validation:

• Confirm or adjust TB and PDC grades for each hotspot.

3. Annotation Refinement:

• Generate final annotations based on one of the annotator’s work by clicking on “final edit” on their name.
• Adjust final annotations as necessary.

4. Final Documentation:

• Final editing of chain of thought for training vision-language models.
• Submit

Rewards

Contributors to this project will be rewarded through a comprehensive system that recognizes both the quality and quantity of their work. Annotators and validators together will receive 70% ownership stakes in the datasets they help create (click on the rightmost information button on each case), which can be appreciated in value as the data is used in research and AI development.

As for this task, we will release this dataset to HuggingFace and make it freely available to the community. Moreover, an AI model will be developed within codatta ecosystem and an online inference service will also be publicly available to all researchers and pathologists worldwide.

References

[1] Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, Bray F. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin. 2021;71(3):209–249.

[2] Zlobec I, Lugli A. Epithelial mesenchymal transition and tumor budding in aggressive colorectal cancer: tumor budding as oncotarget. Oncotarget. 2010;1(7):651–661.

[3] Ueno H, Mochizuki H, Akagi Y, Kusumi T, Yamada K, Ikegami T, et al. Optimal colorectal cancer staging criteria in TNM classification. J Clin Oncol. 2012;30(13):1519–1526.

[4] Lugli A, Kirsch R, Ajioka Y, Bosman F, Cathomas G, Dawson H, et al. Recommendations for reporting tumor budding in colorectal cancer based on the International Tumor Budding Consensus Conference (ITBCC) 2016. Mod Pathol. 2017;30(9):1299–1311.

[5] Rogers AC, Winter DC, Heeney A, Gibbons D, Lugli A, Puppa G, Sheahan K. Systematic review and meta-analysis of the impact of tumour budding in colorectal cancer. Br J Cancer. 2016;115(7):831–840.

[6] Barresi V, Branca G, Ieni A, Reggiani Bonetti L, Baron L, Mondello S, Tuccari G. Poorly differentiated clusters (PDCs) as a novel histological predictor of nodal metastases in pT1 colorectal cancer. Virchows Arch. 2014;464(6):655–662.

[7] Dawson H, Galuppini F, Träger P, Berger MD, Studer P, Brügger L, et al. Validation of the International Tumor Budding Consensus Conference 2016 recommendations on tumor budding in stage I-IV colorectal cancer. Hum Pathol. 2019;85:145–151.

[8] Bosch SL, Teerenstra S, de Wilt JH, Cunningham C, Nagtegaal ID. Predicting lymph node metastasis in pT1 colorectal cancer: a systematic review of risk factors providing rationale for therapy decisions. Endoscopy. 2013;45(10):827–834.

[9] Petrelli F, Tomasello G, Borgonovo K, Ghidini M, Turati L, Dallera P, et al. Prognostic Survival Associated With Left-Sided vs Right-Sided Colon Cancer: A Systematic Review and Meta-analysis. JAMA Oncol. 2017;3(2):211–219.

[10] Zlobec I, Lugli A. Epithelial mesenchymal transition and tumor budding in aggressive colorectal cancer: tumor budding as oncotarget. Oncotarget. 2010;1(7):651–661.

[11] van Wyk HC, Park J, Roxburgh C, Horgan P, Foulis A, McMillan DC. The role of tumour budding in predicting survival in patients with primary operable colorectal cancer: a systematic review. Cancer Treat Rev. 2015;41(2):151–159.

[12] Bokhorst JM, Blank A, Lugli A, Zlobec I, Dawson H, Vieth M, et al. Assessment of individual tumor buds using keratin immunohistochemistry: moderate interobserver agreement suggests a role for machine learning. Mod Pathol. 2020;33(5):825–833.

[13] Puppa G, Senore C, Sheahan K, Vieth M, Lugli A, Zlobec I, et al. Diagnostic reproducibility of tumour budding in colorectal cancer: a multicentre, multinational study using virtual microscopy. Histopathology. 2012;61(4):562–575.

[14] Mitrovic B, Schaeffer DF, Riddell RH, Kirsch R. Tumor budding in colorectal carcinoma: time to take notice. Mod Pathol. 2012;25(10):1315–1325.

[15] Max N, Harbaum L, Pollheimer MJ, Lindtner RA, Kornprat P, Langner C. Tumour budding with and without admixed inflammation: two different sides of the same coin? Br J Cancer. 2016;114(4):368–371.

[16] Banaeeyan R, Fauzi MFA, Chen W, et al. Tumor budding detection in H&E-stained images using deep semantic learning. 2020 IEEE Region 10 Conference (TENCON). IEEE; 2020:52–56.

[17] Bergler M, Benz M, Rauber D, et al. Automatic detection of tumor buds in pan-cytokeratin stained colorectal cancer sections by a hybrid image analysis approach. In: Reyes-Aldasoro CC, Janowczyk A, Veta M, Bankhead P, Sirinukunwattana K, eds. Digital Pathology. Lecture Notes in Computer Science. Springer International Publishing; 2019:83–90.

[18] Bokhorst JM, Rijstenberg L, Goudkade D, Nagtegaal I, van der Laak J, Ciompi F. Automatic detection of tumor budding in colorectal carcinoma with deep learning. In: Stoyanov D, Taylor Z, Ciompi F, et al, eds. Computational Pathology and Ophthalmic Medical Image Analysis. Lecture Notes in Computer Science. Springer International Publishing; 2018:130–138.

[19] Lu J, Liu R, Zhang Y, et al. Development and application of a detection platform for colorectal cancer tumor sprouting pathological characteristics based on artificial intelligence. Intell Med. 2022;2(2):82–87.

[20] Weis CA, Kather JN, Melchers S, et al. Automatic evaluation of tumor budding in immunohistochemically stained colorectal carcinomas and correlation to clinical outcome. Diagn Pathol. 2018;13(1):64.

[21] Fisher NC, Loughrey MB, Coleman HG, Gelbard MD, Bankhead P, Dunne PD. Development of a semi-automated method for tumour budding assessment in colorectal cancer and comparison with manual methods. Histopathology. 2022;80(3):485–500.

[22] Pai RK, Hartman D, Schaeffer DF, et al. Development and initial validation of a deep learning algorithm to quantify histological features in colorectal carcinoma including tumour budding/poorly differentiated clusters. Histopathology. 2021;79(3):391–405.