Transfer Learning-Based Quantification of Mitotic Activity and Its Prognostic and Molecular Associations in Lung Squamous Cell Carcinoma.
5 affiliations
- 1Department of Pulmonology, Bandırma Training and Research Hospital, Bandırma 10200, Türkiye.
- 2Institute of Research and Development, Duy Tan University, Da Nang 550000, Vietnam.
- 3School of Engineering and Technology, Duy Tan University, Da Nang 550000, Vietnam.
- 4Department of Computer Engineering, Bandırma Onyedi Eylül University, Balıkesir 10200, Türkiye.
- 5Engineering and Informatics Department, Klaipėdos Valstybinė Kolegija/Higher Education Institution, 92294 Klaipeda, Lithuania.
Abstract
Background: Automated assessment of mitotic activity may provide a reproducible marker of tumor proliferation, but computational mitosis recognition remains challenging because of cross-domain variability. We developed a transfer-learning framework to quantify mitotic activity in lung squamous cell carcinoma (LUSC) and investigated its clinical and molecular associations. Methods: A ResNet-34 classifier was trained using MIDOG source-domain data and adapted to TCGA-LUSC diagnostic whole-slide images. The target cohort included 478 patients and 512 whole-slide images. Tumor regions and mitotic figures were annotated by two pathologists, and 128 × 128-pixel patches at 40× resolution were analyzed. After identifying patient overlap in the initial patch-level partition, we implemented patient-level stratified five-fold cross-validation to obtain leakage-mitigated performance estimates and out-of-fold predictions. Tumor-area mitosis proportion (TAMP) was defined as the proportion of evaluable tumor-region patches classified as mitosis-positive. Prognostic associations were evaluated using multivariable Cox proportional-hazards regression adjusted for age, sex, and pathological stage. Results: The classifier achieved an AUC of 0.972 (95% CI, 0.965-0.980) on the internal MIDOG test set. Direct application to TCGA-LUSC yielded an AUC of 0.81, whereas target-domain fine-tuning produced an initial patch-level AUC of 0.93. Patient tracing identified overlap in the original partitions, with 142 patients (29.7%) contributing patches to multiple subsets, including 67 patients (14.0%) represented in both training and test sets. Leakage-mitigated patient-level cross-validation yielded a mean AUC of 0.88 (95% CI, 0.86-0.90) and PR-AUC of 0.24. A patient-disjoint WSI holdout evaluation yielded an AUC of 0.87 (95% CI, 0.84-0.90) and PR-AUC of 0.22. High TAMP was associated with poorer overall survival (log-rank p=0.017; HR = 1.73, 95% CI, 1.12-2.68) and remained significant after adjustment for age, sex, and pathological stage (adjusted HR = 1.62, 95% CI, 1.03-2.54; p=0.036). High-TAMP tumors showed enrichment of proliferation- and growth-related transcriptional programs, whereas immune, genomic, and predicted drug-response associations were modest and hypothesis-generating. Conclusions: TAMP represents an image-derived measure of mitotic activity associated with survival and molecular characteristics in TCGA-LUSC. The findings support further investigation of computational mitotic phenotyping, but external multicenter validation and prospective clinical assessment are required before clinical translation.
Related
- Integrative Multi-Database Bioinformatics Characterization of FANCE in Hepatocellular Carcinoma: Prognostic, Epigenetic, Immune, and ncRNA Associations
- Radiomics-based machine learning models decipher heterogeneity of anoikis-related molecular subtypes for prognostic stratification in head and neck squamous cell carcinoma
- SOLUBLE CELL ADHESION MOLECULES QUANTIFICATION IN NEWLY DIAGNOSED DIFFUSE LARGE B-CELL LYMPHOMA PATIENTS: PROGNOSTIC IMPLICATIONS AND CLINICAL-BIOLOGICAL ASSOCIATIONS
- Histology-Stratified Clinical and Prognostic Associations of PDCD-Related Genes in Lung Adenocarcinoma and Lung Squamous Cell Carcinoma.
- Integrated Transcriptomic and Regulatory Network Analysis Reveals Distinct Molecular Subtypes with Prognostic Significance in Clear Cell Renal Cell Carcinoma
BibTeX
@article{gunes2026transfer,
title = {{Transfer Learning-Based Quantification of Mitotic Activity and Its Prognostic and Molecular Associations in Lung Squamous Cell Carcinoma.}},
author = {Gulcin Yilmaz Gunes and Mehmet Akif Cifci},
journal = {Cancers},
year = {2026},
month = sep,
doi = {10.3390/cancers18193098},
url = {https://europepmc.org/article/MED/42882793},
note = {CC BY},
}
RIS
TY - JOUR
TI - Transfer Learning-Based Quantification of Mitotic Activity and Its Prognostic and Molecular Associations in Lung Squamous Cell Carcinoma.
AU - Gulcin Yilmaz Gunes
AU - Mehmet Akif Cifci
PY - 2026
DA - 2026/09/23
JO - Cancers
DO - 10.3390/cancers18193098
UR - https://europepmc.org/article/MED/42882793
AB - Background: Automated assessment of mitotic activity may provide a reproducible marker of tumor proliferation, but computational mitosis recognition remains challenging because of cross-domain variability. We developed a transfer-learning framework to quantify mitotic activity in lung squamous cell carcinoma (LUSC) and investigated its clinical and molecular associations. Methods: A ResNet-34 classifier was trained using MIDOG source-domain data and adapted to TCGA-LUSC diagnostic whole-slide images. The target cohort included 478 patients and 512 whole-slide images. Tumor regions and mitotic figures were annotated by two pathologists, and 128 × 128-pixel patches at 40× resolution were analyzed. After identifying patient overlap in the initial patch-level partition, we implemented patient-level stratified five-fold cross-validation to obtain leakage-mitigated performance estimates and out-of-fold predictions. Tumor-area mitosis proportion (TAMP) was defined as the proportion of evaluable tumor-region patches classified as mitosis-positive. Prognostic associations were evaluated using multivariable Cox proportional-hazards regression adjusted for age, sex, and pathological stage. Results: The classifier achieved an AUC of 0.972 (95% CI, 0.965-0.980) on the internal MIDOG test set. Direct application to TCGA-LUSC yielded an AUC of 0.81, whereas target-domain fine-tuning produced an initial patch-level AUC of 0.93. Patient tracing identified overlap in the original partitions, with 142 patients (29.7%) contributing patches to multiple subsets, including 67 patients (14.0%) represented in both training and test sets. Leakage-mitigated patient-level cross-validation yielded a mean AUC of 0.88 (95% CI, 0.86-0.90) and PR-AUC of 0.24. A patient-disjoint WSI holdout evaluation yielded an AUC of 0.87 (95% CI, 0.84-0.90) and PR-AUC of 0.22. High TAMP was associated with poorer overall survival (log-rank p=0.017; HR = 1.73, 95% CI, 1.12-2.68) and remained significant after adjustment for age, sex, and pathological stage (adjusted HR = 1.62, 95% CI, 1.03-2.54; p=0.036). High-TAMP tumors showed enrichment of proliferation- and growth-related transcriptional programs, whereas immune, genomic, and predicted drug-response associations were modest and hypothesis-generating. Conclusions: TAMP represents an image-derived measure of mitotic activity associated with survival and molecular characteristics in TCGA-LUSC. The findings support further investigation of computational mitotic phenotyping, but external multicenter validation and prospective clinical assessment are required before clinical translation.
SN - 2072-6694
ER -