Abstract
The present study shows that the performance of CNN is not significantly different from the best classical methods and human doctors for classifying mediastinal lymph node metastasis of NSCLC from PET/CT images. Because CNN does not need tumor segmentation or feature calculation, it is more convenient and more objective than the classical methods. However, CNN does not make use of the import diagnostic features, which have been proved more discriminative than the texture features for classifying small-sized lymph nodes. Therefore, incorporating the diagnostic features into CNN is a promising direction for future research.
Introduction
Accurately determining if lung cancer has spread to the mediastinal lymph nodes is critical for deciding between surgery and chemotherapy. Traditional clinical practice often relies on simple thresholds, such as a node size greater than 10mm or a high Standardized Uptake Value (SUV) on PET scans. However, this approach is often subjective and suffers from low sensitivity. While radiomics—the extraction of complex texture features—has been proposed to improve diagnosis, its effectiveness remains a subject of debate. This study investigates whether modern deep learning, specifically Convolutional Neural Networks (CNNs), can provide a more automated and objective alternative to traditional and radiomic-based machine learning methods.
Methods
Researchers conducted a head-to-head comparison of five computational approaches using a retrospective dataset of 1,397 lymph nodes from 168 patients. Every node was pathologically verified as the gold standard.
Computational Approaches
- Classical Machine Learning: Four models—Random Forest (RF), Support Vector Machines (SVM), Adaptive Boosting (AdaBoost), and Artificial Neural Networks (ANN)—were trained on various feature sets:
- D13 (Diagnostic): 13 features used in clinical practice (e.g., size, SUV, CT density).
- T82 (Texture): 82 mathematical descriptors of image heterogeneity.
- A95 & S6: Combined and optimized feature subsets.
- Deep Learning: A modified AlexNet architecture was trained on raw PET and CT image patches. Unlike classical models, the CNN learns discriminative patterns directly from pixels without manual feature engineering.
Key Results
The study found that the choice of features significantly impacts the performance of classical machine learning models. Diagnostic features (D13) consistently outperformed complex texture features (T82).
AdaBoost and RF using diagnostic features emerged as the top classical performers, achieving an Accuracy (ACC) of approximately 85% and an Area Under the Curve (AUC) of 0.91.
The CNN achieved an ACC of 86% and an AUC of 0.91, showing no statistically significant difference from the best-performing classical models. Notably, all machine learning methods demonstrated higher sensitivity than human doctors (84% for CNN vs. 73% for humans), although doctors maintained slightly higher specificity.
Discussion and Conclusion
The results indicate that CNNs can match the performance of hand-crafted radiomic models while offering greater objectivity by eliminating manual segmentation and feature calculation. A surprising finding was the superiority of basic diagnostic features over complex texture features; the authors suggest this is due to the small size of the lymph nodes in the dataset (80% were < 4mm), which provides insufficient voxels for robust texture analysis.
Future work will focus on "feature fusion," integrating raw SUV data directly into CNN architectures to overcome information loss during image normalization. Additionally, multi-center studies are needed to validate these models across diverse patient populations and imaging hardware.
Summary updated