Estimating the Volume of Nodules and Masses on Serial Chest Radiography Using a Deep-Learning-Based Automatic...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging Analysis

Background:

• Pulmonary nodules and masses require accurate volume assessment for effective monitoring.
• Serial chest X-rays (CXRs) are common for follow-up, but precise volumetric analysis can be challenging.

Purpose of the Study:

• To assess pulmonary nodule and mass volumes using parameters derived from a deep-learning-based automatic detection algorithm (DLAD) on serial CXRs.
• To develop and validate a predictive model for nodule volume estimation.

Main Methods:

• A retrospective study included 72 patients with 147 serial CXRs and corresponding CT images.
• A DLAD utilizing a convolutional neural network was developed to detect nodules and derive parameters like area and mean probability from CXRs.
• Volume prediction models were built using regression analyses (linear and non-linear) with DLAD-derived parameters, validated against semi-automatically measured 3D CT volumes.

Main Results:

• The mean nodule/mass volume was 9.37 ± 11.69 cm³.
• DLAD-derived area and CT volume showed moderate linear correlation (R = 0.58).
• Area and mean probability demonstrated a strong linear correlation (R = 0.73), with a multivariable regression model using these parameters achieving the best volume prediction performance (RMSE: 7975.6 mm³).

Conclusions:

• The DLAD-based prediction model, incorporating nodule area and mean probability, offers accurate quantitative estimation of pulmonary nodule/mass volume.
• This approach enables reliable assessment of volume changes on serial CXRs, aiding in patient management.