Integrating CT-based radiomics and deep learning for invasive prediction of ground-glass nodules in lung adenocarcinoma: a multicohort study

Du Hai, Jing Shen, Feng Chen, Kaifeng Wang, Lili Qin, Yijiang Hu, Yue Xiao, Xiulin Wang, Jianlin WuInsights 2025 · DOI

TL;DR

Researchers have developed a hybrid AI framework that combines traditional "hand-crafted" radiomics with deep learning representations to predict the invasiveness of lung nodules from preoperative CT scans. By using a Multiple-Instance Learning (MIL) approach, the model achieved high diagnostic accuracy (AUCs up to 0.926) across multiple hospitals, offering a robust tool to help surgeons decide between conservative lung-sparing surgery and more aggressive lobectomies.

What Problem Does This Paper Address?

Ground-glass nodules (GGNs) are common early signs of lung adenocarcinoma (LUAD). However, not all GGNs are created equal: some are non-invasive (requiring only minor surgery or monitoring), while others are invasive (requiring aggressive surgical resection and lymph node dissection).

The "gold standard" for determining invasiveness is pathology, which can only be done after the nodule is removed. Preoperatively, radiologists rely on CT scans, but manual assessment is subjective and prone to error. While previous AI attempts have used either radiomics (extracting quantitative features like shape and texture) or deep learning (letting neural networks find their own patterns), both have flaws. Radiomics is highly dependent on precise segmentation, while deep learning is often a "black box" that can struggle with the 3D complexity of medical images. This paper seeks to bridge that gap by fusing both methods into a single, robust framework.

Methods

The authors proposed a Multiple-Instance Learning (MIL) framework called MIL-DL-Rad. Here’s how the "sausage is made":

  1. Feature Extraction: The system extracts three types of data:
  2. Radiomics: 1,834 hand-crafted features describing the nodule's shape, intensity, and texture.
  3. 2.5D Deep Learning: Features from several 2D slices around the nodule's center.
  4. 3D Deep Learning: Volumetric features captured by 3D convolutional neural networks.
  5. The MIL Framework: Instead of treating a nodule as a single image, MIL treats it as a "bag" of instances (slices). This allows the model to weigh the importance of different parts of the nodule, mimicking how a radiologist scrolls through a CT stack.
  6. Feature Fusion & Selection: Using $t$-tests, Pearson correlation, and LASSO regression, the team narrowed down the massive feature set to the 49 most predictive markers.
  7. Classification: They tested several machine learning classifiers, finding that ExtraTrees (Extremely Randomized Trees) provided the most stable and accurate results.

Datasets

This was a large-scale, multicenter study involving 1,247 GGNs from 1,182 patients across six different hospitals. The data was partitioned to ensure the model wasn't just "memorizing" one hospital's imaging style:

  • Training & Validation Sets: Sourced from two hospitals ($n=289$ and $n=542$).
  • External Test Sets: Three independent cohorts from four other hospitals ($n=219, 116, \text{and } 81$) were used to prove the model works in the "real world" on different CT scanners.

Key Insights and Ideas

  • The Power of Fusion: The "aha moment" was that radiomics and deep learning are complementary. Radiomics captures macro-level morphology (like how "spiky" or "round" a nodule is), while deep learning picks up on subtle, abstract textures that the human eye (and manual formulas) might miss.
  • Stability via MIL: Single-modality models (just DL or just radiomics) showed "unstable" performance across different test sets. The MIL-DL-Rad model remained consistent, suggesting that the MIL framework effectively filters out the "noise" inherent in medical imaging.
  • Clinical Utility: Beyond just accuracy, the authors used Decision Curve Analysis (DCA) to show that using this model provides a higher "net benefit" in clinical decision-making compared to treating all patients or treating none, specifically in the threshold ranges where surgical decisions are most difficult.

Key Results

The integrated MIL-DL-Rad model outperformed all other variations (Radiomics-only, 2.5D-only, 3D-only, and MIL-DL without radiomics).

  • AUC Performance: The model achieved an Area Under the Curve (AUC) of 0.936 in the training set.
  • Generalizability: More importantly, it maintained high performance on external test sets with AUCs of 0.868, 0.926, and 0.918.
  • Comparison: In the validation set, the integrated model (AUC 0.881) significantly outperformed the 3D deep learning model (AUC 0.802) and the pure MIL-DL model (AUC 0.778).

Limitations and Future Work

Despite the impressive numbers, the authors highlight a few hurdles:

  • Manual Segmentation: The nodules were manually delineated by radiologists, which is time-consuming. Future iterations should incorporate automatic segmentation to speed up the workflow.
  • Retrospective Bias: The study used historical data. A prospective, randomized controlled trial is needed to see how this tool performs when used by surgeons in real-time.
  • Data Imbalance: The training sets were relatively small compared to the total population, and there was an imbalance between invasive and non-invasive cases, which can sometimes skew AI learning.
  • "Black Box" Elements: While radiomics adds interpretability, the deep learning components remain difficult to explain fully, which can be a barrier to clinical adoption.

Summary updated

Abstract

Ground-glass nodules invasiveness directly influences surgical strategies and prognosis. Multiple-instance learning framework integrates radiomics and deep learning features. Integrated model achieves superior accuracy and consistency in invasiveness prediction.
BibTeX
@inproceedings{hai2025integrating,
  author = {Du Hai and Jing Shen and Feng Chen and Kaifeng Wang and Lili Qin and Yijiang Hu and Yue Xiao and Xiulin Wang and Jianlin Wu},
  title = {{Integrating CT-based radiomics and deep learning for invasive prediction of ground-glass nodules in lung adenocarcinoma: a multicohort study}},
  booktitle = {Insights},
  year = {2025},
  doi = {10.1186/s13244-025-02156-6},
}
Figure