Room: AAPM ePoster Library
Purpose: in vendor implementation of Exposure Index (EI) have been widely observed in digital radiography (DR), especially in detector area used for EI calculation in clinical images. Determination of EI based on anatomy of interest can minimize variances caused by other anatomies, foreign objects or raw radiation. This study aims to investigate the feasibility of anatomy-specific EI (EI-A) in lateral lumbar spine radiographs and compare it with a vendor proprietary EI calculation.
Methods: “for-processing” lateral lumbar spine images were de-identified and collected from multiple DR systems of a vendor. First, EI calculation based on raw pixel values and detector sensitivity was confirmed. Second, lumbar spine was segmented in each image using a semi-automated approach involving image contrast adjustment, edge detection, and human intervention and verification using MATLAB. These images were further augmented by shifting, smoothing, and adding Gaussian noise to generate 560 images. Data was used to train a convolutional neural network for automated segmentation based on the U-Net architecture, with 70% for training, 20% for validation, and 10% for testing. Next, pixel values from the lumbar spines were used to compute EI-A through three methods, i.e., median value, first quantile, and first peak from pixel histogram multi-peak Gaussian fitting. These values were compared with the vendor calculated EI values.
Results: accuracy of the convolutional neural network reached 94.6%. EI-A values were different compared to vendor EI values to various degrees. The median pixel value method yielded -42.4% to 15.3% differences relative to the current, the first quantile method yielded -58.2% to 6.4%, while the first peak method yielded -64.0% to -9.3% differences .
Conclusion: of utilizing anatomy-specific pixel values to calculate EI has been demonstrated in clinical lateral lumbar spine images. Three methods were compared to emphasize different aspects of detector air kerma corresponding to the spine area.