Machine learning links immune-cell measures to resected lung cancer prognosis

Machine-learning analysis of whole-slide tissue images found that CD8-positive tumor-infiltrating lymphocytes, epithelial Foxp3-positive lymphocytes, and PD-L1 expression were associated with survival in a retrospective cohort of 165 patients who underwent surgery for non-small cell lung cancer. The association was particularly evident among patients with stage I disease, for whom high CD8-positive lymphocyte density remained independently associated with longer disease-free survival. The study assessed immune features in epithelium and stroma, separating the two tissue compartments and then evaluating them together, and showed that quantitative digital pathology supplied prognostic information in addition to conventional pathological stage. These were observational associations from one Japanese cancer center, not evidence that measuring or targeting any of the features would improve treatment.

The investigators reviewed patients who had surgery at Shizuoka Cancer Center between September 2014 and September 2015. Eligible cases included stage I to stage IIIA non-small cell lung cancer, while patients receiving adjuvant platinum-based chemotherapy or immunotherapy were excluded. That restriction was intended to reduce the influence of postoperative systemic treatment on the assessment of immune variables. The final cohort was relatively small and centered on early-stage disease: 75.9% of tumors were pathological stage I and 75.2% were adenocarcinomas. Median follow-up for patients alive at their last visit was 60 months, with a reported range of 7 to 74 months. Disease-free survival was the primary endpoint, measured from surgery to recurrence or death, while overall survival was a secondary endpoint measured to death from any cause or the last follow-up visit.

To perform the digital analysis, the researchers selected formalin-fixed, paraffin-embedded surgical tissue blocks containing each tumor’s center and invasive margin. Slides were scanned at high resolution, and the investigators used HALO image-analysis software, whose tissue-classification system is based on a Random Forest machine-learning algorithm. The system used an ensemble of eight random forests, each containing 10 trees, with predictions averaged across the ensemble. Each forest used four default features: two color features based on Gaussian analysis and two texture features based on Hessian-of-Gaussian and structure-tensor analysis. Pathologists manually annotated tumor areas on hematoxylin-and-eosin-stained whole-slide images and excluded large necrotic regions where those areas could not be distinguished reliably from epithelium on AE1/AE3 pancytokeratin staining.

The same tissue classifier then separated epithelial, stromal, and non-epithelial/non-stromal regions by comparing the whole-slide images of the H&E and AE1/AE3-stained sections. The last category mainly included minor necrosis and mucus, and its lymphocytes were not included in the analysis. Separate CD8 and Foxp3 slides were processed with pathologist-selected parameters for nuclear segmentation and staining-intensity thresholds. The researchers then used the HALO multiplex immunohistochemistry algorithm to count CD8-positive and Foxp3-positive tumor-infiltrating lymphocytes separately in the epithelial and stromal compartments. Densities were reported as positive cells per square millimeter, rather than as counts from selected high-power fields. The PD-L1 tumor proportion score was calculated as the percentage of positive malignant cells among all malignant cells in the epithelial region.

Measurements were made across the whole annotated tumor region in one operation. The investigators also combined epithelial and stromal areas, but emphasized that this combined measurement did not represent the entire tumor because the amount of unevaluated necrosis differed among specimens. Across the combined regions, median CD8-positive density was 432.33 cells per square millimeter, compared with 335.59 cells per square millimeter in epithelium and 582.06 in stroma. Median Foxp3-positive density was 123.66 cells per square millimeter in the combined regions, with medians of 73.23 in epithelium and 179.20 in stroma. Median PD-L1 tumor proportion score was 16.27%, and values ranged from 0.093% to 77.45%. Epithelial and stromal CD8-positive densities were strongly correlated, with a Pearson coefficient of 0.81 and p<0.001; the corresponding Foxp3-positive correlation was 0.52, also with p<0.001. High CD8-positive density was associated with adenocarcinoma at p=0.011, while most other clinicopathological characteristics showed no significant relationship with the measured immune factors.

For survival comparisons, patients were divided at the cohort median for each immune measurement. Higher CD8-positive density was associated with better disease-free and overall survival whether the density was measured in epithelium, stroma, or both regions. For disease-free survival, the log-rank test yielded p<0.001 for the combined and epithelial CD8 measurements and p=0.003 for the stromal measurement. In multivariable Cox models, high combined CD8-positive density was independently associated with lower disease-free-survival hazard, with a hazard ratio of 0.43 and 95% confidence interval of 0.26 to 0.72. High PD-L1 tumor proportion score was also independently associated with disease-free survival, with a hazard ratio of 0.49 and 95% confidence interval of 0.30 to 0.81, as was high combined Foxp3-positive density, with a hazard ratio of 0.55 and 95% confidence interval of 0.34 to 0.89. Pathological stage remained an independent factor in each model; stage II or III disease had a disease-free-survival hazard ratio of 3.65 to 4.09 relative to stage I, depending on the immune variable in the model.

The overall-survival analysis produced a more selective pattern. High CD8-positive density remained independently associated with lower mortality hazard, with a hazard ratio of 0.37 and 95% confidence interval of 0.19 to 0.72, and high PD-L1 score did as well, with a hazard ratio of 0.48 and 95% confidence interval of 0.25 to 0.90. The association for combined Foxp3-positive density was not statistically significant in the multivariable analysis: its hazard ratio was 0.61, with a 95% confidence interval of 0.33 to 1.143 and p=0.11. Compartment-specific, unadjusted analyses showed that Foxp3-positive density in epithelium was associated with better disease-free and overall survival, whereas stromal Foxp3-positive density was not significantly associated with either endpoint. Thus, the location of that immune-cell measurement affected the observed results, while the study’s multivariable analysis did not establish Foxp3-positive density as an independent overall-survival marker.

Stage-specific analyses indicated that the digital immune measurements could distinguish prognosis among patients whose cancers shared the same broad TNM category. Among stage I patients, five-year disease-free survival was 81.5% for those with high combined CD8-positive density and 57.6% for those with low density, a difference with p=0.003. The corresponding figures for high and low combined Foxp3-positive density were 81.3% and 59.1%, with p=0.035. Stage I patients with high PD-L1 score had a five-year disease-free survival of 80.8%, compared with 60.3% among those with a low score, although that difference did not cross the study’s statistical-significance threshold of p<0.05, with p=0.053. For overall survival in stage I disease, the high-versus-low five-year estimates were 90.9% versus 70.8% for CD8-positive density and 90.3% versus 72.9% for PD-L1 score. In multivariable analysis limited to stage I patients, high CD8-positive density was independently associated with disease-free survival, with a hazard ratio of 0.39 and 95% confidence interval of 0.20 to 0.77.

The findings come from a retrospective, single-institution cohort and have several limitations. The sample was relatively small, and the analysis did not examine pre-existing lung conditions such as emphysema or interstitial lung disease in detail. It evaluated only three features in a tumor microenvironment composed of many cell types. Bias may also have affected the pathologist’s development of the algorithm, and the manual exclusion of substantial necrosis meant that the epithelial and stromal areas did not equal the whole tumor. The use of median-derived cutoffs limits what can be inferred about specific thresholds, while the selected Japanese population and exclusion of patients receiving adjuvant treatment restrict how broadly the results can be applied. The study therefore establishes associations and the feasibility of automated immune quantification in surgical specimens; it does not demonstrate that the measurements caused better outcomes or determine which postoperative treatment would benefit an individual patient.

Despite those constraints, the work shows how computational pathology can make spatial immune measurements across a resection specimen more systematic than manual counting in selected microscopic fields. CD8-positive density had the most consistent relationship with outcomes across the analyses, while epithelial and stromal CD8 measurements were strongly correlated and each was associated with survival. Measuring the two compartments together also retained prognostic associations and could simplify analysis, although the study did not establish that this approach provided a statistically stronger separation than separate assessment. The authors noted that more advanced deep-learning systems might distinguish tumor and nontumor regions automatically and reduce manual annotation. For the present evidence, the central result is that machine-learning-based whole-slide analysis identified reproducible, spatially resolved immune measures that supplied prognostic information beyond pathological stage, while their role in postoperative treatment selection remains to be established.

Subject of Research: Retrospective prognostic analysis of digitally measured tumor immune features in surgically resected non-small cell lung cancer

Article Title: Prognostic significance of tumor microenvironment assessed by machine learning algorithm in surgically resected non‐small cell lung cancer

Article References: Terada, Y., Isaka, M., Ono, A., Kawata, T., Serizawa, M., Mori, K., Muramatsu, K., Tone, K., Kenmotsu, H., Ohshima, K., Urakami, K., Nagashima, T., Kusuhara, M., Akiyama, Y., Sugino, T., Takahashi, T., & Ohde, Y. (2026). Prognostic significance of tumor microenvironment assessed by machine learning algorithm in surgically resected non‐small cell lung cancer. Cancer Reports, 9(10), Article e1926. https://doi.org/10.1002/cnr2.1926

Image Credits: AI Generated

DOI: 10.1002/cnr2.1926

Keywords: non-small cell lung cancer, machine learning, digital pathology, whole-slide imaging, tumor-infiltrating lymphocytes, CD8-positive T cells, Foxp3-positive T cells, PD-L1, cancer prognosis, computational pathology, Prognostic, significance