Deep learning based prediction of solid component infiltration in lung adenocarcinoma under different CT thresholds(PDF)
《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]
- Issue:
- 2026年第1期
- Page:
- 60-64
- Research Field:
- 医学影像物理
- Publishing date:
Info
- Title:
- Deep learning based prediction of solid component infiltration in lung adenocarcinoma under different CT thresholds
- Author(s):
- CHEN Aide; HU Yaheng; WEI Wencun
- Department of Equipment, Ankang Hospital of Traditional Chinese Medicine, Ankang 725000, China
- Keywords:
- Keywords: lung adenocarcinoma solid component lung nodule deep learning CT threshold
- PACS:
- R318；R734.2
- DOI:
- DOI:10.3969/j.issn.1005-202X.2026.01.008
- Abstract:
- Abstract: Objective To explore the effects of different CT thresholds on the performance of 3 deep learning models in lung nodule classification tasks for identifying the optimal threshold that enhances model classification performance and thereby providing a reference for the accurate diagnosis of lung nodules. Methods CT imaging data were collected from 113 patients with lung adenocarcinoma (involving 176 lung nodules) in a Grade A tertiary hospital. Five different CT thresholds were selected, and the lung nodules were classified using convolutional neural network, residual network, and convolutional long short-term memory network. The area under the receiver operating characteristic curve (AUC-ROC), accuracy, precision, and recall of each model under different thresholds were calculated. Results The classification performance of all 3 models generally improved as the CT threshold gradually increased. Especially at the threshold of -200 HU, the AUC-ROC values of all models reached their peaks, which were 0.855 for convolutional neural network, 0.870 for residual network, and 0.860 for convolutional long short-term memory network. However, when the threshold was further increased to -100 HU, the model performance declined. Conclusion The selection of CT thresholds exerts significant effects on the performance of deep learning models in lung nodule classification. The threshold of -200 HU may be the optimal choice as it yields the best model performance. Future computer-aided diagnostic systems should consider performing CT image preprocessing near this threshold to improve the diagnostic accuracy of lung nodules.
Last Update: 2026-01-27