Exploring the Relevance of Full Lung CT Scans in Deep Learning for Classifying Nodule Malignancy in Lung Cancer Screening
Keywords
Loading...
Authors
Issue Date
2024-06-14
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
1) Background: Current screening guidelines for radiologists
in lung cancer state the relevance of examining not only nodule
features in a CT scan, but also other abnormalities before
classifying a nodule on malignancy risk. In deep learning models
in literature, some models focus only on nodule features while
others also take larger parts of the CT scan into account.
However, due to their difference in model characteristics, it is
hard to know which features are important to analyse.
2) Purpose: To develop a deep learning algorithm that can
analyse the added value of outside-the-nodule (global) features
on top of a nodule (local) feature focused model.
3) Materials and Methods: A two-stream deep learning
algorithm was trained, where one stream analyses nodule block
features and the other stream analyses features outside the
nodule block. These parts are then combined to create a final
nodule malignancy risk estimation. The model was trained with
nodules from National Lung Screening Trial and validated
with screening data from Danish Lung Cancer Screening Trial,
and Multicentric Italian Lung Detection trial, and clinical data
from Radboud UMC. From each testset, a nodule size-matched
subset was gathered. Furthermore, the impact of screening
variables emphysema and nodule location were analysed.
Other experiments for model interpretability were performed
using uncertainty cross-examinations and pixel attribution maps.
4) Results: For the test-data from screening trial DLCST
the global, the local, and the combined model reached an AUC
of 0.85, 0.93, and 0.93 respectively. For the screening trial of
MILD, the models reached an AUC of 0.81, 0.97, 0.96 AUC. For
the clinical data, the models had an AUC of 0.68, 0.90 and 0.89.
Size matching benign and malignant nodules led to an overall
decrease of performance for all models across all datasets.
The screening variable nodule location showed significant
influence on the outside-the-nodule malignancy risk estimations
in both DLCST and MILD (p <0.001). The screening variable
emphysema, which was only available in DLCST, did not have
a significant influence (p = 0.091).
Examination of uncertainty showed a decrease of each model
performance on its own uncertain subset. Cross-examination
showed a decrease in global model performance, when the
local model was uncertain. However, if the global model was
uncertain, the local model still performed well.
Finally, Pixel-attribution examination shows the global model
focuses on lung structure and bone tissue.
5) Conclusion: We managed to build the two-stream model
and it converged to a reasonable performance. The combined
model did not improve upon the local model, yet the global model
performed above expectations without access to nodule characteristics.
Furthermore, the global model verified the significance
of Brock model variable nodule location for screening. Finally,
pixel-attribution maps show focus on bone-tissue features, which
needs further investigation.
Index Terms—Lung cancer screening, Pulmonary nodule, Lung
CT Scans, Deep learning, Medical Imaging, Machine learning,
Emphysema, Upper lobe nodule, Bone-tissue, Coronary calcification,
Brock model, Explainable AI, Pixel-attribution maps,
Uncertainty
Description
Citation
Faculty
Faculteit der Sociale Wetenschappen