Computed Tomography-based Prognostication in Lung Adenocarcinomas through Histopathological Feature Learning: A
Kyung Hee Lee1,2, Jong Hyuk Lee3, Samina Park4
1Department of Radiology and.
Annals of the American Thoracic Society
|April 19, 2023
Summary
Deep learning models using computed tomography (CT) scans predict survival in early-stage lung adenocarcinoma by learning histopathological features. This CT-based composite score demonstrates high reproducibility and complements existing clinical factors for improved prognostication.
Area of Science:
- Radiology
- Oncology
- Artificial Intelligence
Background:
- Accurate prognostication in early-stage lung adenocarcinomas is crucial for treatment planning.
- Histopathological risk factors are well-validated but require invasive procedures.
- Imaging surrogates could offer non-invasive prognostic tools.
Purpose of the Study:
- To develop and validate computed tomography (CT)-based deep learning (DL) models for predicting survival in early-stage lung adenocarcinomas.
- To assess the models' ability to learn histopathological features from CT scans.
- To investigate the reproducibility of these DL models using multicenter data.
Main Methods:
- Two DL models were trained on preoperative CT scans from 1,426 patients to predict visceral pleural invasion and lymphovascular invasion.
- A composite score was derived from the averaged model outputs.
- The composite score's prognostic discrimination and added value were evaluated in temporal and external test sets of stage I lung adenocarcinomas, assessing freedom from recurrence (FFR) and overall survival (OS).
- Interscan and interreader reproducibility were analyzed.
Main Results:
- The CT-based composite score showed good prognostic discrimination for 5-year FFR (AUC 0.76) and 5-year OS (AUC 0.67-0.69) in test sets.
- The performance remained stable over a 10-year follow-up.
- The composite score provided independent and complementary prognostic value to clinicopathological factors (P < 0.001).
- Excellent interscan and interreader reproducibility (Pearson's correlation coefficient, 0.98) were observed.
Conclusions:
- CT-based deep learning models can effectively learn histopathological features for prognostication in early-stage lung adenocarcinomas.
- The developed composite score demonstrates significant prognostic value and high reproducibility.
- This approach offers a promising non-invasive tool for improving survival prediction in lung cancer.
Keywords:
artificial intelligencelung cancermultidetector computed tomographyprediction modelprognosis
