# Improving early liver metastasis detection in colorectal cancer using a weighted ensemble of ResNet50 and swin transformer: a KHCC study

## TL;DR

This study develops an AI model combining ResNet50 and Swin Transformer to detect early liver metastasis in colorectal cancer patients, improving detection accuracy and speed.

## Contribution

A novel weighted ensemble deep learning model for early detection of liver metastasis in colorectal cancer using architectural diversity.

## Key findings

- The ensemble model achieved 75.48% accuracy and 0.8115 AUC in detecting liver metastasis.
- The model outperformed individual architectures with statistically significant improvements.
- The framework is computationally efficient, requiring only 0.39 seconds per image inference.

## Abstract

Colorectal cancer represents the third most diagnosed malignancy globally, with liver metastasis occurring in approximately 50–60% of patients following initial treatment. Current surveillance strategies utilizing carcinoembryonic antigen monitoring and interval cross-sectional imaging demonstrate significant limitations in early hepatic recurrence detection, often identifying disease at advanced, unresectable stages. This study addresses the critical research gap in AI-driven surveillance frameworks by developing a novel ensemble deep learning model for early liver metastasis prediction in colorectal cancer patients. The methodology employed six state-of-the-art architectures including ResNet50, MobileNetV2, DenseNet121, CNN-LSTM, and Swin Transformer as feature extractors through transfer learning, followed by weighted soft voting ensemble learning combining the top-performing models. The framework was evaluated on a comprehensive dataset of 1,628 medical images from colorectal cancer patients, with rigorous statistical validation using Friedman and Wilcoxon signed-rank tests. Results demonstrated that the ensemble model combining ResNet50 and Swin Transformer achieved superior performance with 75.48% accuracy, 79.0% sensitivity, 73.6% specificity, and 0.8115 AUC, representing statistically significant improvements over all individual architectures. The ensemble approach successfully addressed the challenging nature of the dataset where multiple state-of-the-art models achieved near-random performance, demonstrating the effectiveness of architectural diversity in medical image analysis. The clinical impact of this work extends to enhancing early detection capabilities that could increase patient eligibility for curative interventions, with balanced diagnostic performance suitable for surveillance applications. The computationally efficient framework requires only 0.39 s per image inference time, making it feasible for integration into existing clinical workflows and potentially improving outcomes for colorectal cancer patients through earlier identification of hepatic recurrence.