Quantitative assessment of lung nodule detectability using pixel value-based receiver operating characteristics analysis

BackgroundOptimizing operational protocols in medical imaging is essential to ensure the quality of radiological diagnoses. However, a quantitative method for evaluating the image quality of actual patients and detectability of lesions within these clinical images has not yet been established.PurposeTo quantitatively assess the difficulty in detecting nodules on chest radiographs using a pixel value (PV)-based receiver operating characteristic (ROC) analysis approach.Material and MethodsA chest radiograph database from the Japanese Society of Radiological Technology-containing lung nodule images classified into five levels of detection difficulty-was used for analysis. Multiple regions of interest (ROIs) were defined to encompass both nodules and surrounding anatomical structures. The mean PV and standard deviation values were calculated for each region. Assuming normal PV distributions for both nodules and backgrounds, the PV-based area under the ROC curve (AUC) was computed using a theoretical formula. The method's validity was verified by analyzing correlations with the subtlety classification, which reflects detection difficulty.ResultsAnalysis of 154 nodule images demonstrated a strong correlation with nodule subtlety (r = 0.998), and with observer-derived AUC values (r = 0.955), confirming the effectiveness of the proposed metric.ConclusionThe proposed method enables quantitative evaluation of lesion detectability in clinical images. This novel index may offer valuable clinical feedback for optimizing imaging conditions and can serve as a practical tool for training in diagnostic radiology.