A dual-branch CNN-LSTM framework for exploratory classification of colorectal cancer histopathology and hepatocellular carcinoma clinical records
- 1Department of Infectious Diseases, The Third Affiliated Hospital of Sun Yat-sen University , , ,
- 2Department of Medical Oncology, The Sixth Affiliated Hospital of Sun-Yat Sen University , , ,
Abstract
Colorectal cancer (CRC) and hepatocellular carcinoma (HCC) remain major diagnostic challenges because classification often depends on modality-specific evidence that is stored in separate public datasets. This study evaluates a dual-branch CNN-LSTM framework using two independent and unpaired data sources: colorectal histopathological image patches and structured HCC clinical-biochemical records. The study is therefore positioned as an exploratory investigation of representation learning across heterogeneous medical datasets. The CNN branch extracts spatial and glandular texture features from curated CRC image patches, whereas the LSTM branch treats fixed-order clinical attributes as a structured feature sequence to learn inter-feature dependencies in the HCC records. Under internal public-dataset testing with bootstrap uncertainty estimation, the full CNN + LSTM + MLP framework achieved an accuracy of 0.978 (95% CI, 0.971-0.984), Macro-F1 of 0.977 (95% CI, 0.971-0.984), and Macro OVR AUC of 0.997 (95% CI, 0.996-0.998). Ablation and baseline comparisons showed lower performance for CNN only (accuracy 0.938), LSTM only (0.875), clinical encoder only (0.890), CNN + LSTM (0.966), CNN + clinical MLP (0.958), late fusion (0.952), EfficientNet-B0-style fusion (0.970), ViT-tiny-style fusion (0.960), XGBoost (0.910), random forest (0.900), and logistic regression (0.875). The full model also exceeded the EfficientNet-B0-style baseline in Macro AUC by 0.003 (95% CI, 0.002-0.004; paired bootstrap P = 0.007; McNemar exact P = 0.011). Grad-CAM and SHAP analyses are used to interpret branch-specific decisions, while the findings remain limited to internal public-dataset validation.
Related
- EffAtt-LungNet: An attention-guided EfficientNet framework for three-class lung histopathology classification
- Calibration-Aware Patient-Level Prediction of Source-Defined MSIMUT Status in Colorectal Cancer from Whole-Slide Histopathology Images Using Foundation-Model Embeddings
- Histopathology image classification based on ensembles of highly compact features
- An exploratory study of circulating cell-free mRNAs across colorectal cancer stages
- CT Radiomics and Machine Learning for Three-Class Bone Density Classification: An Exploratory Comparison Across Lumbar Vertebral Levels
BibTeX
@article{guo2026dualbranch,
title = {{A dual-branch CNN-LSTM framework for exploratory classification of colorectal cancer histopathology and hepatocellular carcinoma clinical records}},
author = {Xiaoyan Guo and Jing Lai and Lili Wu and Dianke Chen},
journal = {Frontiers in oncology},
year = {2026},
month = sep,
doi = {10.3389/fonc.2026.1878061},
url = {https://europepmc.org/article/PMC/PMC13661494},
note = {CC BY},
}
RIS
TY - JOUR
TI - A dual-branch CNN-LSTM framework for exploratory classification of colorectal cancer histopathology and hepatocellular carcinoma clinical records
AU - Xiaoyan Guo
AU - Jing Lai
AU - Lili Wu
AU - Dianke Chen
PY - 2026
DA - 2026/09/30
JO - Frontiers in oncology
DO - 10.3389/fonc.2026.1878061
UR - https://europepmc.org/article/PMC/PMC13661494
AB - Colorectal cancer (CRC) and hepatocellular carcinoma (HCC) remain major diagnostic challenges because classification often depends on modality-specific evidence that is stored in separate public datasets. This study evaluates a dual-branch CNN-LSTM framework using two independent and unpaired data sources: colorectal histopathological image patches and structured HCC clinical-biochemical records. The study is therefore positioned as an exploratory investigation of representation learning across heterogeneous medical datasets. The CNN branch extracts spatial and glandular texture features from curated CRC image patches, whereas the LSTM branch treats fixed-order clinical attributes as a structured feature sequence to learn inter-feature dependencies in the HCC records. Under internal public-dataset testing with bootstrap uncertainty estimation, the full CNN + LSTM + MLP framework achieved an accuracy of 0.978 (95% CI, 0.971-0.984), Macro-F1 of 0.977 (95% CI, 0.971-0.984), and Macro OVR AUC of 0.997 (95% CI, 0.996-0.998). Ablation and baseline comparisons showed lower performance for CNN only (accuracy 0.938), LSTM only (0.875), clinical encoder only (0.890), CNN + LSTM (0.966), CNN + clinical MLP (0.958), late fusion (0.952), EfficientNet-B0-style fusion (0.970), ViT-tiny-style fusion (0.960), XGBoost (0.910), random forest (0.900), and logistic regression (0.875). The full model also exceeded the EfficientNet-B0-style baseline in Macro AUC by 0.003 (95% CI, 0.002-0.004; paired bootstrap P = 0.007; McNemar exact P = 0.011). Grad-CAM and SHAP analyses are used to interpret branch-specific decisions, while the findings remain limited to internal public-dataset validation.
SN - 2234-943X
ER -