Establishment of a Model to Predict the Prognosis of Endometrial Carcinoma Using Tumor-Infiltrating Lymphocytes Evaluated With Artificial Intelligence: A Retrospective Analysis.

Cancer Rep (Hoboken) 2026 AI 6 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
The Immune System's Role in Endometrial Cancer

Endometrial carcinoma (EC) is the most common gynecological malignancy worldwide, accounting for 4.3% of cancers in women. Treatment typically involves surgery (total hysterectomy with bilateral salpingo-oophorectomy), followed by adjuvant radiation or chemotherapy based on recurrence risk factors.

Tumor-infiltrating lymphocytes (TILs) are immune cells that migrate from the bloodstream into tumor tissue. The density and pattern of TIL infiltration reflects the immune system's recognition and response to the tumor, and has been shown to be a prognostic marker in several cancer types including breast, colon, and ovarian cancers.

In EC, TIL status has been found to correlate with important clinical features including FIGO stage, myometrial invasion (how deeply the tumor invades the uterine wall), lymph node metastasis, and lymphovascular invasion. However, standard TIL evaluation relies on manual counting by pathologists - a process that is time-consuming and subject to significant observer-to-observer variability.

TIL status is also closely connected to mismatch repair (MMR) deficiency, one of the molecular classification pillars of EC. MMR-deficient tumors tend to accumulate many mutations, producing neoantigens that attract immune cells, leading to higher TIL density. This biological link makes AI-based TIL quantification potentially valuable for risk stratification.

TL;DR: Tumor-infiltrating lymphocytes are a prognostic biomarker in endometrial cancer that are currently assessed manually by pathologists - introducing AI-based quantification could standardize this important measurement.
Pages 2-3
AI-Based Lymphocyte Counting on Pathology Slides

This retrospective study included 659 patients with endometrial carcinoma who underwent primary surgery at the National Defense Medical College Hospital in Japan between 1989 and 2022 - one of the largest single-institution series of its kind.

For each patient, one hematoxylin and eosin (H&E) stained slide - the standard pathology stain used in routine diagnosis - was selected from the most invasive frontier of the tumor. This is the area where tumor cells are actively pushing into surrounding tissue, a biologically important zone for immune activity.

The AI system was trained to automatically detect lymphocytes within a defined area: a 500 micrometer-wide band extending 250 micrometers toward the stroma (connective tissue surrounding the tumor) and 250 micrometers toward the tumor from the invasive frontline. This precise definition was selected based on prior literature and exploratory analyses comparing different candidate zones.

The AI classifier's accuracy was validated using the F1 score (the harmonic mean of precision and recall), comparing AI-detected lymphocytes against manual expert annotations. An F1 score of 0.81 was achieved - within the range of 0.78-0.97 reported in similar TIL AI studies, confirming acceptable accuracy.

TL;DR: An AI system automatically counted lymphocytes in a precisely defined zone around the invasive tumor frontier on standard H&E slides, achieving F1 score of 0.81 against expert manual annotation.
Page 3
Defining High-TIL and Low-TIL Groups

After AI quantification, the average number of lymphocytes per unit area (per micrometer squared) was calculated for all 659 patients. The optimal cutoff value separating High-TIL from Low-TIL groups was determined using ROC curve analysis for predicting cancer recurrence or progression.

The area under the ROC curve was 0.570 and the cutoff value of 0.00478 lymphocytes per micrometer squared was selected. Sensitivity analyses using alternative cutoffs (first quartile, median, third quartile) showed similar prognostic trends, confirming the result was robust to cutoff selection.

Using this threshold, patients were classified into the High-TIL group (346 patients, 52.5%) and the Low-TIL group (313 patients, 47.5%). The near-equal split suggests the cutoff captures a genuine biological distribution rather than an extreme outlier boundary.

Immunohistochemical staining for four MMR proteins (MLH1, MSH2, MSH6, and PMS2) was performed to assess MMR status, allowing the study to explore the relationship between TIL density and the mismatch repair deficiency molecular subtype of EC.

TL;DR: Patients were divided into High-TIL and Low-TIL groups using an ROC-derived cutoff, with nearly equal group sizes, and MMR protein staining was performed to explore the biological connection.
Pages 5-6
High TILs Predict Better Survival

Patients in the High-TIL group had significantly better progression-free survival (PFS) - the time from surgery until disease recurrence or death - than those in the Low-TIL group (p < 0.01). Similarly, overall survival (OS) was significantly better in the High-TIL group (p < 0.01).

On multivariate analysis (controlling for other prognostic factors like tumor grade, stage, and myometrial invasion), TIL status remained an independent prognostic factor: High-TIL status reduced the risk of recurrence or progression by 39% (Hazard Ratio 0.61, 95% CI 0.43-0.87) and the risk of death by 46% (Hazard Ratio 0.54, 95% CI 0.33-0.86).

The High-TIL group was diagnosed at earlier FIGO stages and had a higher frequency of myometrial invasion less than 50% - findings consistent with prior studies that evaluated TILs manually, confirming that AI-based quantification captures the same biological signal as expert human counting.

MMR deficiency (dMMR) was significantly more common in the High-TIL group than in the Low-TIL group (p < 0.01), supporting the known biological link between mismatch repair deficiency and heightened immune infiltration through neoantigen-driven immune activation.

TL;DR: High TIL density was independently associated with significantly better progression-free and overall survival in endometrial cancer, reducing risk of recurrence by 39% and risk of death by 46%.
Pages 5-6
AI TIL Evaluation Validates Prior Human Studies

A key strength of this study is that the AI-based findings align consistently with results from previous studies using manual TIL assessment. The same associations between TIL density and FIGO stage, myometrial invasion, and MMR status were observed, providing independent confirmation that AI quantification is biologically valid.

The study also confirmed the association between TILs and dMMR in EC. Given that approximately 25%-30% of EC patients have dMMR tumors, and these patients may benefit from immunotherapy (immune checkpoint inhibitors), AI-based TIL quantification could potentially serve as an accessible surrogate marker for immunotherapy response.

An important unresolved question is how TIL assessment should be standardized across different EC molecular subtypes. The prognostic value of TILs may differ between POLE-mutated, MSI-high, copy-number-low, and p53-abnormal subtypes, requiring subgroup-specific analyses in future studies.

The study evaluated TILs using H&E staining, which cannot distinguish between different lymphocyte subtypes (CD4+, CD8+, regulatory T cells). Future work should examine whether the clinical relevance of TILs differs depending on the specific immune cell type identified by AI.

TL;DR: AI-based TIL quantification validates manual pathology findings and opens the door to standardized, reproducible immune assessment that could inform immunotherapy decisions in endometrial cancer.
Page 6
Toward Standardized AI Pathology in Endometrial Cancer

This study demonstrates that AI-based quantification of tumor-infiltrating lymphocytes on standard H&E pathology slides can accurately and significantly predict prognosis in endometrial cancer patients - without requiring any additional staining or molecular testing.

The prognostic value is independent of established clinical risk factors, meaning TIL assessment adds information beyond what tumor grade, stage, and histological subtype alone provide. This could help clinicians better identify which early-stage patients are at higher risk of recurrence and need more intensive surveillance or adjuvant treatment.

The AI approach addresses a fundamental limitation of manual TIL evaluation: it is objective, reproducible, and scalable. Unlike human counting, which varies between observers and is impractical for routine large-scale use, an automated system can deliver consistent results across any pathology laboratory.

Future work should integrate AI-based TIL assessment with complete molecular classification (including POLE status) in larger prospective cohorts, and explore whether this measurement can guide selection of patients most likely to respond to immune checkpoint inhibitors such as pembrolizumab.

TL;DR: AI-based TIL quantification from standard pathology slides provides independent prognostic information in endometrial cancer, offering a scalable, reproducible tool that could improve risk stratification and treatment decisions.
Citation: Open Access, 2026. Available at: PMC13066499.