Related Experiment Video
Updated: Aug 1, 2025
07:53
Author Spotlight: Advancing 3D Modeling for Enhanced Diagnosis and Treatment of Pulmonary Nodules in Early-Stage Lung Cancer
Published on: October 13, 2023
1.5K
Using 3D deep features from CT scans for cancer prognosis based on a video classification model: A multi-dataset
Junhua Chen1, Leonard Wee1, Andre Dekker1
1Department of Radiation Oncology (MAASTRO), GROW School for Oncology and Developmental Biology, Maastricht University Medical Centre+, Maastricht, Netherlands.
Medical Physics
|April 27, 2023
Summary
Deep learning imaging features show promise in cancer prognosis, outperforming traditional radiomics in predictive accuracy. However, radiomics currently offers better reproducibility and interpretability for clinical decision-making.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Accurate cancer prognosis is crucial for patient management.
- Handcrafted imaging biomarkers (radiomics) show potential in predicting patient outcomes.
Purpose of the Study:
- To investigate if deep learning-based 3D imaging features can serve as effective imaging biomarkers.
- To compare the performance of deep learning features against radiomics in cancer prognosis.
Main Methods:
- Deep learning features were extracted using a pre-trained Inflated 3D ConvNet (I3D) architecture on CT scans transformed into videos.
- Radiomics served as the reference imaging biomarker.
- The study utilized four diverse datasets (LUNG 1, LUNG 4, OPC, H&N 1) comprising 1270 cancer samples.
Main Results:
- Deep learning features demonstrated superior predictive accuracy for survival prediction compared to radiomics across multiple datasets (e.g., CI of 0.87 vs. 0.77 in LUNG 4).
- A significant portion of selected deep features were not correlated with tumor volume or TNM staging.
- Radiomics features exhibited higher reproducibility in a test/retest setting (CCC 0.89) than deep features (CCC 0.62).
Conclusions:
- Deep learning imaging features can outperform radiomics in predicting cancer prognosis.
- Deep features offer novel insights into tumor prognosis beyond traditional clinical factors.
- Challenges remain regarding the reproducibility and interpretability of deep learning features compared to radiomics.
