CellScope: Optimising Papillary Thyroid Cancer Diagnosis

Transcript: Papillary thyroid carcinoma, or PTC, is the most common type of thyroid cancer. While many cases have an excellent prognosis, some variants, especially Tall Cell and Columnar Cell, are more aggressive and may need more extensive treatment and closer follow-up. The challenge is that these aggressive forms are still diagnosed by visually estimating whether tumour cells are at least three times taller than they are wide. In borderline cases, that can be hard to judge consistently. Here, you can see the difference: Classic PTC cells are shorter and less elongated, while tall cell and columnar cell tumours have taller and narrower cells. To make this process more objective, I used InstaSeg in QuPath to measure individual tumour cells on digital slides. I trained CellScope, my multivariable screening model, on data from 307 cases. When using the tool, the CSV file exported from InstaSeg is entered along with genetic and clinical data to generate a screening result.

Why:

In the United States and Canada, about 52,140 thyroid cancer cases are expected this year1,2. Papillary Thyroid Carcinoma (PTC) makes up about 80-85% of thyroid cancers3. While many PTCs have excellent outcomes, aggressive variants such as Tall Cell Variant (TCV) and Columnar Cell Variant (CCV) have a worse prognosis, higher recurrence risk and need more aggressive treatment4,5. In a 2025 propensity-matched study of 1,065 patients, TCV had a 5-year disease-free survival of 48.5%, compared with 73.1% in classical PTC5. (Figure 1)

TCV is defined mainly by appearance under the microscope. WHO 2022 criteria requires at least 30% of tumour cells to be at-least three times taller than wide6. Because pathologists cannot measure every cell on a slide, borderline cases depend on visual estimation, creating interobserver variability. This project explores whether AI can reduce that subjectivity by measuring cell shape and whether genotype leaves a visible morphologic difference. BRAF V600E is the most common driver mutation in PTC, occurring in about 45-50% of cases, while RAS mutations occur in about 10-20% commonly occuring in follicular-pattern tumours7,8.

Goal:

Develop an AI-based screening approach that measures tumour cell length-to-width ratios and combines morphometry with mutation data to help identify aggressive PTC subtypes more objectively and efficiently.

Hypothesis:

If AI-based cell measurements are combined with genetic mutation data, aggressive PTC subtypes can be screened more accurately, objectively, and efficiently than by visual estimation alone because these tumours show detectable differences in cell morphology and mutation patterns.

Background Research:

Searched peer-reviewed thyroid cancer studies, including PubMed, and the public TCGA-THCA thyroid carcinoma project from the NIH Genomic Data Commons9.

Variables:

Constants: TCGA-THCA cohort, WHO Tall Cell criteria, ICD-O 8344/3 grouping for CCV+TCV.

Independent: subtype, BRAF and RAS mutation status, AJCC stage, age.

Dependent: cell L:W ratio, model probability of aggressive variant.

Data Collection:

Used the public TCGA-THCA dataset from the peer-reviewed NIH Genomic Data Commons9. 307 cases: 242 Classic PTC, 37 Tall Cell and Columnar Cell (ICD-O 8344/3), 28 Follicular Variant, with whole-slide images, BRAF/RAS calls, age, gender, stage, progression-free survival, and treatment.

AI Cell Segmentation:

Ran InstaSeg10 inside QuPath11, scripted in Groovy, to outline tumour cells and measure their maximum and minimum diameters. Calculated L:W per cell for 20–30 cells per case.

Cross-Platform Check:

Owkin Pathology Explorer12, a separate AI pathology deep-learning tool, measured cell circularity to independently confirm elongation differences between subtypes were not a QuPath result.

CellScope Screening Model:

Ridge-regularised multivariable logistic regression13 with six inputs: L:W, BRAF, RAS, AJCC stage, age, and an L:W × BRAF interaction term.

Statistical Testing:

I used p-values, the Kruskal–Wallis test14, and ROC/AUC analysis15. I also used repeated 5×5-fold cross-validation16 and 2,000-times bootstrap confidence intervals17 to test reliability.

Manual Validation:

In a random subset of 20 cases, I manually measured tumour cells under pathologist supervision and compared those results with the AI measurements.

Experiment:

Finally, two pathologists independently timed and diagnosed the same TCV case. I compared their review times and diagnoses with the AI multivariable model. Likert-scale survey18 was used for qualitative feedback. (Figure 2)

Controls: same slide, computer, stopwatch, room.

Independent Variable: Pathologist 1 and 2, InstaSeg/CellScope

Dependent Variable: Diagnostic time, diagnostic accuracy

Experimental Errors: human timing error, differences in pathologist experience/interpretation, observer bias.

CellScope identified aggressive papillary thyroid carcinoma (PTC) cases in the TCGA-THCA dataset9 with an AUC of 0.859 on 307 cases. AUC is a score from 0.5 (random guessing) to 1.0 (perfect separation)15. My multivariable model was stronger than using cell shape alone, which gave an AUC of 0.769. (Figure 3)

1. Cell shape differed by subtype

The median length-to-width (L:W) ratio was highest in the aggressive CCV/TCV group at 1.702, compared with 1.621 in Classic PTC and 1.582 in Follicular Variant.

The differences were statistically significant for:

CCV vs Classic PTC: p = 1 x 10-6

CCV vs Follicular Variant: p = 1.2 x 10-5

Classic PTC vs Follicular Variant was not significant since they are both less aggressive subtypes and showed similar L:W distributions.

When all three groups were compared together using the Kruskal–Wallis test, the result was also highly significant:

H = 31.17, p = 1.7 x 10-7

This test was used because it compares three groups at once. These results show that aggressive tumour cells are measurably more elongated. (Figure 4)

2. Genetics also changed cell shape

Tumours with a BRAF mutation had higher L:W ratios than tumours with no BRAF mutation and higher L:W ratios than RAS-mutated tumours.

BRAF-mutated vs no mutation: p = 0.000149

BRAF-mutated vs RAS-mutated: p = 0.000027

This means BRAF V600E tumour cells tended to be taller and narrower. In other words, genetics left a visible difference in morphology (cell shape) that the AI could detect. (Figure 5)

3. New finding inside Classic PTC

The most interesting result appeared inside the Classic PTC group itself. BRAF V600E cases had a higher mean L:W ratio than BRAF-wildtype (Which is the gene without the mutation)19 cases (1.644 vs 1.620, p = 0.018). They also had a higher L:W ratio than RAS cases (p = 0.002).

This suggests that a tall-cell-like shape may already be appearing in some tumours still labelled classic.

4. Screening performance

At the rule-out threshold, sensitivity was 0.87 and NPV was 0.97. This means the model caught 87% of aggressive cases and, when it cleared a case, it was correct 97% of the time.

At the more balanced threshold, sensitivity was 0.76, specificity was 0.85, PPV was 0.41, and NPV was 0.96. This means it still caught most aggressive cases, correctly ruled out 85% of non-aggressive cases, and a flagged case was 3.4 times more likely to be aggressive than a random case in this dataset.

5. Pathologist vs AI

Pathologist 1: 9 min 6 sec and diagnosed TCV.

Pathologist 2: 9 min 56 sec and diagnosed TCV.

InstaSeg AI: 3 min 50 sec and flagged TCV.

Likert scale survey18 results in Figure 6.

6. Manual validation

In 20 cases reviewed under pathologist supervision, my manual L:W measurements closely matched InstaSeg’s AI measurements. This showed that the tool produced accurate and trustworthy ratios.

Unlike previous AI studies in PTC, which mainly focused on detecting tall-cell areas or generating tall-cell scores20,21, CellScope measures the length-to-width ratio of each individual tumour cell and uses those measurements combined with genetic/clinical information. This turns the WHO tall-cell definition6 from a subjective visual estimation into a quantitative measurement, therefore decreasing interobserver variability and improving consistency.

The importance of these findings is shown by their strong statistical significance. The cell-shape differences across subtypes were highly significant (Kruskal–Wallis H = 31.17, p = 1.7 x 10-7), meaning they are extremely unlikely to be due to chance. The full model also performed strongly, with an AUC of 0.859 across 25 cross-validation rounds, showing good separation between aggressive and non-aggressive cases.

The practical value of this result is strongest in borderline cases, where CellScope gives pathologists quantitative evidence instead of relying on visual estimation. It also saves time. In my experiment, both pathologists and the AI reached the same TCV diagnosis, but CellScope finished in 3 min 50 sec of machine time, compared with 9–10 minutes for visual review. Because that is machine time, the pathologist can spend those minutes on other cases, reviewing consultations, or writing reports while the program runs. On a larger scale, that could save specialist time, lower labour cost per screened case, and help reduce backlog in a field already facing pressure in Canada22.

Next, I plan to focus on validating CellScope on larger external datasets, especially on cohorts with more Tall Cell and Columnar Cell cases. I also want to link CellScope directly to InstaSeg so that the length to width ratio is extracted instantly from a slide. Another next step is expanding the pathologist comparison to more cases and reviewers. Longer term, with more validation and pathologist feedback, I hope to make CellScope an open access and practical screening tool for researchers and pathology labs.

I would like to sincerely thank everyone who helped and supported me throughout my project. Firstly, I would like to thank my parents and my sister for always encouraging and believing in me every step of this journey.  I also extend my thanks to my mentor, Dr. Matthew Cecchini, for his advice and mentorship. Special thanks to all the participants who took part in my experiment. Each of you played an essential role in my project and I am truly thankful for your help.

1) American Cancer Society. (2026, January 13). Key statistics for thyroid cancer.

https://www.cancer.org/cancer/types/thyroid-cancer/about/key-statistics.html

2) Canadian Cancer Society. (2026). Thyroid cancer statistics.

https://cancer.ca/en/cancer-information/cancer-types/thyroid/statistics

3) Limaiem, F., & Nassereddine, S. (2024). Papillary thyroid carcinoma. In StatPearls. StatPearls Publishing.

https://www.ncbi.nlm.nih.gov/books/NBK536943

4) Coca-Pelaz, A., Shah, J. P., Hernandez-Prera, J. C., Ghossein, R. A., Rodrigo, J. P., Hartl, D. M., Olsen, K. D., Shaha, A. R., Zafereo, M., Suarez, C., Nixon, I. J., Randolph, G. W., Mäkitie, A. A., Kowalski, L. P., Vander Poorten, V., Sanabria, A., Guntinas-Lichius, O., Simo, R., Zbären, P., Angelos, P., Khafif, A., Rinaldo, A., & Ferlito, A. (2020). Papillary thyroid cancer—Aggressive variants and impact on management: A narrative review. Advances in Therapy, 37(7), 3112–3128. doi:10.1007/s12325-020-01391-1

https://pubmed.ncbi.nlm.nih.gov/32488657

5) Parvathareddy, S. K., Siraj, A. K., Qadri, Z., Al-Rasheed, M., Haqawi, W., Siraj, N., Al-Sobhi, S. S., Al-Dayel, F., & Al-Kuraya, K. S. (2025). Tall cell variant histology predicts poorer disease-free survival in papillary thyroid carcinoma: A propensity-matched cohort study. World Journal of Surgery, 49(9), 2443–2448. doi:10.1002/wjs.12683

https://pubmed.ncbi.nlm.nih.gov/40714960/

6) Turchini, J., et al. (2023). A critical assessment of diagnostic criteria for the tall cell variant of papillary thyroid carcinoma. Endocrine Pathology, 35, 79–88.

https://www.researchgate.net/publication/404031712_Establishment_and_validation_of_a_nomogram_model_for_poorly_differentiated_thyroid_cancer_based_on_the_Surveillance_Epidemiology_and_End_Result_database

7) Riesco-Eizaguirre, G., et al. (2025). BRAF V600E in thyroid cancer: Navigating prognostic controversies and therapeutic opportunities. Frontiers in Endocrinology, 16, Article 1635516.

https://pmc.ncbi.nlm.nih.gov/articles/PMC12741945/

8) Gupta, N., Dasyam, A. K., & Carty, S. E. (2013). RAS mutations in thyroid FNA specimens are highly predictive of predominantly low-risk follicular-pattern cancers. The Journal of Clinical Endocrinology & Metabolism, 98(5), E914–E922.

https://pmc.ncbi.nlm.nih.gov/articles/PMC5393462/

9) National Cancer Institute Genomic Data Commons. (n.d.). TCGA-THCA. GDC Data Portal. https://portal.gdc.cancer.gov/projects/TCGA-THCA

10) QuPath docs authors. (2019–2025). InstanSeg. In QuPath 0.6.0 documentation. https://qupath.readthedocs.io/en/0.6/docs/deep/instanseg.html

11) Bankhead, P., Loughrey, M. B., Fernández, J. A., Dombrowski, Y., McArt, D. G., Dunne, P. D., McQuaid, S., Gray, R. T., Murray, L. J., Coleman, H. G., James, J. A., Salto-Tellez, M., & Hamilton, P. W. (2017). QuPath: Open source software for digital pathology image analysis. Scientific Reports, 7, 16878. doi:10.1038/s41598-017-17204-5

https://www.nature.com/articles/s41598-017-17204-5

12) Owkin. (2026, February 9). Case study: Pathology Explorer.

https://www.owkin.com/blogs-case-studies/case-study-pathology-explorer

13) Friedrich, S., Groll, A., Ickstadt, K., Kneib, T., Pauly, M., Rahnenführer, J., & Friede, T. (2023). Regularization approaches in clinical biostatistics: A review of methods and their applications. Statistical Methods in Medical Research, 32(2), 425–440. doi:10.1177/09622802221133557

https://pubmed.ncbi.nlm.nih.gov/36384320/

14) National Institute of Standards and Technology. (n.d.). Kruskal-Wallis test.

https://www.itl.nist.gov/div898/software/dataplot/refman1/auxillar/kruskwal.htm

15) Nahm, F. S. (2022). Receiver operating characteristic curve: Overview and practical use for clinicians. Korean Journal of Anesthesiology, 75(1), 25–36.

pubmed.ncbi.nlm.nih.gov/35124947/

16) Bradshaw, T. J., Huemann, Z. M., & Rahmim, A. (2023). A guide to cross-validation for artificial intelligence in medical imaging. Radiology: Artificial Intelligence, 5(4), e220232.

https://pmc.ncbi.nlm.nih.gov/articles/PMC10388213/

17) National Institute of Standards and Technology. (n.d.). Bootstrap plot.

https://www.itl.nist.gov/div898/handbook/eda/section3/bootplot.htm

18) McLeod, S. (2025, January 31). Likert scale questionnaire: Meaning, Examples & Analysis. Simply Psychology.

https://www.simplypsychology.org/likert-scale.html

19) National Cancer Institute. (n.d.). Definition of wild-type gene.

https://www.cancer.gov/publications/dictionaries/cancer-terms/def/wild-type-gene

20) Stenman, S., Linder, N., Lundin, M., Haglund, C., Arola, J., & Lundin, J. (2022). A deep learning–based algorithm for tall cell detection in papillary thyroid carcinoma. PLOS ONE, 17(8), e0272696. doi:10.1371/journal.pone.0272696

https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0272696&

21) Stenman, S., et al. (2024). External validation of a deep learning-based algorithm for detection of tall cells in papillary thyroid carcinoma: A multicenter study. Journal of Pathology Informatics, 15, 100366. doi:10.1016/j.jpi.2024.100366

https://www.sciencedirect.com/science/article/pii/S2153353924000051

22) College of American Pathologists–Association canadienne des pathologistes. (2025, April 11). Message from the President.

https://cap-acp.org/news/698547/Message-From-The-President.htm

Image/Other References:

MyEndoConsult. (n.d.). Histology of papillary thyroid carcinoma. My Endo Consult. https://myendoconsult.com/learn/histology-of-papillary-thyroid-carcinoma

PathologyOutlines.com. (n.d.). Thyroid gland & parathyroid: Tall cell variant of papillary thyroid carcinoma. PathologyOutlines.com. https://www.pathologyoutlines.com/topic/thyroidtallcellvariant.html

QuPath docs authors. (2019–2026). InstanSeg. In QuPath 0.6.0 documentation. Read the Docs. https://qupath.readthedocs.io/en/0.6/docs/deep/instanseg.html

QuPath docs authors. (2019–2026). Welcome to QuPath!. In QuPath 0.6.0 documentation. Read the Docs. https://qupath.readthedocs.io/en/0.6/

National Cancer Institute Genomic Data Commons. (n.d.). TCGA-THCA. GDC Data Portal. https://portal.gdc.cancer.gov/projects/TCGA-THCA