Leveraging Serial Low-Dose CT Scans in Radiomics-based Reinforcement Learning to Improve Early Diagnosis of Lung Cancer at Baseline Screening

Abstract

Purpose To evaluate the feasibility of leveraging serial low-dose CT (LDCT) scans to develop a radiomics-based reinforcement learning (RRL) model for improving early diagnosis of lung cancer at baseline screening. Materials and Methods In this retrospective study, 1951 participants (female patients, 822; median age, 61 years [range, 55-74 years]) (male patients, 1129; median age, 62 years [range, 55-74 years]) were randomly selected from the National Lung Screening Trial between August 2002 and April 2004. An RRL model using serial LDCT scans (S-RRL) was trained and validated using data from 1404 participants (372 with lung cancer) containing 2525 available serial LDCT scans up to 3 years. A baseline RRL (B-RRL) model was trained with only LDCT scans acquired at baseline screening for comparison. The 547 held-out individuals (150 with lung cancer) were used as an independent test set for performance evaluation. The area under the receiver operating characteristic curve (AUC) and the net reclassification index (NRI) were used to assess the performances of the models in the classification of screen-detected nodules. Results Deployment to the held-out baseline scans showed that the S-RRL model achieved a significantly higher test AUC (0.88 [95% CI: 0.85, 0.91]) than both the Brock model (AUC, 0.84 [95% CI: 0.81, 0.88]; P = .02) and the B-RRL model (AUC, 0.86 [95% CI: 0.83, 0.90]; P = .02). Lung cancer risk stratification was significantly improved by the S-RRL model as compared with Lung CT Screening Reporting and Data System (NRI, 0.29; P < .001) and the Brock model (NRI, 0.12; P = .008). Conclusion The S-RRL model demonstrated the potential to improve early diagnosis and risk stratification for lung cancer at baseline screening as compared with the B-RRL model and clinical models. Keywords: Radiomics-based Reinforcement Learning, Lung Cancer Screening, Low-Dose CT, Machine Learning © RSNA, 2024 Supplemental material is available for this article.

Keywords: Low-Dose CT; Lung Cancer Screening; Machine Learning; Radiomics-based Reinforcement Learning.

Graphical abstract

Figure 1:

The flowchart shows the number of participants diagnosed as positive, negative, or indeterminate for lung cancer at each screening year (T0, T1, and T2), and the splitting of the training/validation and test data sets for model development and evaluation. LDCT = low-dose CT, NLST = National Lung Screening Trial.

Figure 2:

Test receiver operating characteristic curves for classification of lung cancer by the radiomics-based reinforcement learning (RRL) models and the Brock model when deployed to the low-dose CT (LDCT) scans at (A) the baseline screening examinations and (B) the diagnosis year examinations. B-RRL = baseline-year radiomics-based reinforcement learning, D-RRL = diagnosis-year radiomics-based reinforcement learning, S-RRL = serial-year radiomics-based reinforcement learning.

Figure 3:

Axial low-dose CT images show examples of nodules without contrast media that were classified as Lung CT Screening Reporting and Data System (Lung-RADS) 3 or 4A by the serial-year radiomics-based reinforcement learning (S-RRL) model and the Brock model at the time of baseline examinations. (A) With the baseline scan, the S-RRL model correctly identified two benign nodules (underwent 2 years of follow-up scans) as low risk, while the Brock model identified them as medium risk. (B) A benign nodule was identified as medium and high risk by the S-RRL and Brock models, respectively. (C) Two malignant nodules were identified as high risk by the S-RRL model, but the Brock model identified them as low risk. (D) A malignant nodule was identified as medium risk by both models.

Figure 4:

Axial low-dose CT images show examples of small nodules (5 mm or less) without contrast media classified by the serial-year radiomics-based reinforcement learning (S-RRL) model and the Brock model at the baseline screening year. (A) Two malignant nodules were mistakenly classified as low risk by both models (false negatives). (B) The Brock model correctly classified this nodule as low risk. It was the only false-positive classification by the S-RRL model, likely because of the potential of nodule growth predicted by the S-RRL model which was confirmed at the follow-up scans.

Figure 5:

Axial low-dose CT images show examples of nodules without contrast media with different margins and internal characteristics classified by the serial-year radiomics-based reinforcement learning (S-RRL) model and the Brock model at baseline screening examination. (A) The S-RRL model correctly diagnosed three benign nodules as low risk (underwent 2 years of follow-up scans) (true negatives), while the Brock model mistakenly identified them as high or medium risk (false positives). (B) Three benign nodules were mistakenly identified as high risk (false positives) by both models. (C) Three malignant nodules were correctly diagnosed as high risk by the S-RRL model (true positives), but the Brock model diagnosed them as medium risk (false negatives). (D) Three malignant nodules were mistakenly identified as medium risk by both models (false negatives). GGO = ground-glass opacity.

Figure 6:

Bar graph shows the distribution of reclassified individuals (A) without cancer and (B) with lung cancer in different nodule characteristic categories. GGO = ground-glass opacity, Lung_RADS = Lung CT Screening Reporting and Data System.