CT-DeepLearning Team Research: 2.5D Residual Networks Outperform in Pulmonary Nodule Malignancy-Risk Classification

Summary:CT-DeepLearning Team's research, published on arXiv, compares 2D, 2.5D, and 3D residual networks for pulmonary nodule malignancy-risk classification. The study evaluates momentum-contrast pretraining, mild augmentation, and the exclusion of indeterminate ratings. Results show that the 2.5D model, enhanced with data augmentation, achieves the highest performance with a Macro-F1 of 0.6537, outperforming 2D (0.6134) and 3D (0.5403) models. Data augmentation consistently improved performance across

---

Background and Objectives

Malignancy-risk classification of pulmonary nodules is a critical task in medical image analysis, crucial for early lung cancer screening. Traditional methods often rely on 2D or 3D convolutional neural networks (CNNs), but the performance of models with different dimensions in processing CT data varies significantly. This study aims to systematically compare 2D, 2.5D, and 3D residual networks and evaluate the effects of momentum-contrast pretraining and mild data augmentation.

Methods

The study used the LIDC-IDRI dataset, containing 2,654 pulmonary nodules labeled as benign, indeterminate, or malignant. The dataset was split into 1,876 training samples, 382 validation samples, and 396 test samples. Model inputs included a single axial slice, five neighboring slices, and a local CT volume. The experiments were divided into three main parts:

1. 2D Model: Using a single axial slice as input.
2. 2.5D Model: Using five neighboring slices as input.
3. 3D Model: Using a local CT volume as input.

Each model was evaluated with three training methods: random initialization baseline, momentum-contrast pretraining followed by fine-tuning, and supervised training with mild data augmentation.

Key Findings

1. Performance Comparison:

   • The enhanced 2.5D model achieved the highest Macro-F1 score of 0.6537 on the test set, significantly outperforming the 2D (0.6134) and 3D (0.5403) models.
   • The 2D model had a Macro-F1 of 0.6134, and the 3D model had a Macro-F1 of 0.5403.
   • Data augmentation consistently improved performance across all dimensions, while momentum-contrast pretraining did not yield the expected benefits.

2. Binary Classification Task:

   • In the binary classification task, excluding indeterminate labels, the 2.5D model had the highest Macro-F1 of 0.8664.
   • The 2D and 3D models had Macro-F1 scores of 0.8227 and 0.7554, respectively.

3. Conclusion:

   • The 2.5D model performed best in pulmonary nodule malignancy-risk classification, with data augmentation being a key factor in improving model performance.
   • Pretraining methods did not significantly enhance model performance, suggesting the need for further exploration of other pretraining strategies.

Engineering Trade-offs and Recommendations

1. Dimension Selection: The 2.5D model strikes a good balance between performance and computational cost, making it suitable for resource-constrained environments.
2. Data Augmentation: Mild data augmentation is an effective method to enhance model robustness and is recommended for similar tasks.
3. Pretraining Strategy: Momentum-contrast pretraining did not show significant benefits in this task, and exploring other pretraining methods or adjusting the pretraining strategy may be necessary.

Developer Deployment Recommendations

1. Model Selection: It is recommended to prioritize the 2.5D model combined with data augmentation to achieve efficient and high-performance pulmonary nodule classification.
2. Computational Resources: The 2.5D model has moderate computational resource requirements and is suitable for deployment on medium-sized GPU clusters.
3. Continuous Optimization: Further exploration of other pretraining methods and data augmentation strategies is recommended to further enhance model performance.

---

Source: ArXiv Machine Learning (cs.LG) (2026-10-28)

— END —

Tags: #Medical AI Follow #Deep Learning Follow #Residual Networks Follow #Data Augmentation Follow #Pulmonary Nodule Classification Follow