# A deep learning model to enhance lung cancer detection using ‘Dual-Branch’ model classification approach

## TL;DR

This paper introduces a new deep learning model called DbMCA that improves lung cancer detection using a dual-input approach, achieving high accuracy even with limited data.

## Contribution

The novel Dual-Branch Model Classification Approach (DbMCA) combines image and mask data to enhance lung cancer detection accuracy and scalability.

## Key findings

- DbMCA achieved 91.21% accuracy and 91.18% F1-score on a smaller dataset and 98.04% accuracy and 98.01% F1-score on a larger dataset.
- DNN and SVM outperformed CNN on sparse mask data, highlighting architecture sensitivity to sparsity.
- DbMCA showed improved discriminative power for detecting subtle lung cancer patterns but faces limitations like high computational costs and dataset biases.

## Abstract

Cancer remains a life-threatening global challenge, with lung cancer ranking among the most devastating forms, impacting millions annually. Early detection and accurate classification are essential for improving patient survival rates, and computed tomography (CT) has become a critical tool in lung cancer diagnosis. Despite advancements, previous studies have faced notable challenges, particularly a shortage of available samples and limitations in input modalities, both of which hinder model performance. Addressing these issues, this research introduces the Dual-Branch Model Classification Approach (DbMCA), a two-stage strategy that integrates image and mask data to enhance detection accuracy and scalability. Two comparative experiments were conducted using the LIDC-IDRI dataset with varying data sizes to evaluate the impact of sample size and dual-input modalities. The DbMCA achieved remarkable results, as it performed higher accuracy results a 91.21% accuracy and 91.18% F1-score in the smaller dataset and an exceptional 98.04% accuracy and 98.01% F1-score in the larger dataset. CNN performance on sparse mask data declines with scale, while DNN and SVM consistently outperform it, highlighting architecture sensitivity to sparsity. This demonstrates the model’s improved discriminative power and potential for detecting subtle lung cancer patterns, however, based on statistical evidence DbMCA significantly outperforms weaker baselines and successfully integrates multi-modal information. Nonetheless, certain limitations were observed, such as the high computational requirements stemming from large sample sizes, the constrained information provided by segmentation masks, and the presence of potential biases in the dataset. These challenges hinder the model’s ability to generalize effectively. Future research should aim to enhance image quality, broaden the scope of datasets, and overcome segmentation-related constraints to make further progress in lung cancer detection. The DbMCA represents a significant step forward in improving the performance and scalability of diagnostic tools, offering the potential for more effective and lifesaving interventions in lung cancer care.