CC-DCNet: Dynamic Convolutional Neural Network with Contrastive Constraints for Identifying Lung Cancer Subtypes on Multi-modality Images

Yuan Jin, Gege Ma, Geng Chen, Tianling Lyu, Jan Egger, Junhui Lyu, Shaoting Zhang, Wentao ZhuarXiv 2024 · DOI · arXiv

Introduction

Accurately distinguishing between lung cancer subtypes, specifically Lung Adenocarcinoma (LUAD) and Lung Squamous Cell Carcinoma (LUSC), is a critical step in clinical oncology. These subtypes necessitate distinct therapeutic strategies and have different prognostic outcomes. While histopathological examination remains the gold standard for diagnosis, it is inherently invasive and may be unfeasible for patients in advanced stages or with poor physical status.

Computed Tomography (CT) offers a non-invasive alternative; however, the macroscopic nature of CT scans often lacks the resolution required to identify subtle histological features, leading to high inter-observer variability even among expert radiologists. Existing deep learning approaches often struggle with the "modality gap"—the difficulty in aligning high-resolution pathological features with lower-resolution radiological data—and the practical reality of missing data in clinical settings.

Methodology: The CC-DCNet Architecture

To address these challenges, we present CC-DCNet, a Dynamic Convolutional Neural Network with Contrastive Constraints. This framework is designed to leverage the rich information in pathological images to "teach" a radiological model, while remaining flexible enough to handle incomplete datasets.

The CC-DCNet pipeline showing how CT and pathological features are extracted and fused.

The architecture incorporates three primary innovations:

  1. Dynamic Convolution (DC): Traditional static networks are limited by fixed parameters. CC-DCNet utilizes parallel convolution kernels that are dynamically assembled based on the input modality. This allows the model to process paired (CT and pathology) or standalone CT data seamlessly.
  2. Vision Transformer (ViT) Backbones: The model utilizes 3D ViTs for CT volumes and 2D ViTs for pathological patches. A key modification for pathology is the use of magnification encoding instead of standard position encoding, allowing the model to interpret tissue features relative to their biological scale.
  3. The "Or Gate" Mechanism: This logical component allows the training process to incorporate both paired and unpaired data, maximizing the utility of diverse clinical datasets.

Bridging the Modality Gap with Contrastive Constraints

A central innovation of CC-DCNet is the Contrastive Constraint Module. During the training phase, a contrastive loss function is employed to map radiological and pathological features into a shared latent space.

Diagram of the contrastive loss, showing how LUAD and LUSC features are clustered while cross-modality features are aligned.

By minimizing the distance between features of the same subtype across different modalities, the model forces the CT feature extractor to identify "radiomic" signatures that correlate with microscopic tissue patterns. Essentially, pathology acts as a structural guide, enabling the model to "see" microscopic indicators within macroscopic CT scans.

Experimental Results

The model was evaluated using a large-scale multi-center dataset of 1,146 patients from three distinct hospitals. This diverse cohort ensured the model's robustness against variations in imaging equipment and clinical protocols. CC-DCNet demonstrated state-of-the-art performance across multiple testing scenarios:

  • CT-Only Testing: In the most common clinical scenario where only CT scans are available, the model achieved an Accuracy (ACC) of 86.69% and an Area Under the Curve (AUC) of 90.75%.
  • Multi-Modality Testing: When both CT and pathology data were provided, performance reached an ACC of 96.28% and an AUC of 97.53%.
  • Ablation Studies: Comparative tests confirmed that the removal of contrastive constraints led to a significant performance decline, validating the necessity of cross-modality learning.

Conclusion and Future Directions

CC-DCNet represents a significant advancement in non-invasive lung cancer subtyping. By effectively bridging the gap between radiological and pathological data through dynamic convolutions and contrastive learning, it provides a powerful tool for clinical decision support.

Future work will focus on enhancing the interpretability of the model to visualize the specific CT regions linked to pathological features. Additionally, we aim to extend this framework to incorporate other data modalities, such as genomics and clinical metadata, to further refine diagnostic accuracy and generalizability across global healthcare settings.

Summary updated

BibTeX
@misc{jin2024cc,
  author = {Yuan Jin and Gege Ma and Geng Chen and Tianling Lyu and Jan Egger and Junhui Lyu and Shaoting Zhang and Wentao Zhu},
  title = {{CC-DCNet: Dynamic Convolutional Neural Network with Contrastive Constraints for Identifying Lung Cancer Subtypes on Multi-modality Images}},
  year = {2024},
  eprint = {2407.13092},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2407.13092},
}
Figure