Fractal-dimension-based quantification of CT micronodules for classifying borderline pneumoconiosis.

BACKGROUND

Distinguishing category 0/1 from 1/0-1/1 on CT is subject to reader variability. We developed a method combining micronodule count with fractal analysis.

METHODS

We analyzed three-dimensional chest CT images from 48 individuals with coal workers' pneumoconiosis or silicosis, read by consensus using ICOERD. Lung lobes were segmented with software and micronodules manually segmented. For each lobe, we computed micronodule count and fractal dimension from micronodule masks. Features were aggregated as the most affected lobe, defined by highest micronodule count, and as a count-weighted whole-lung average. A linear-kernel support vector machine was evaluated with 5-fold stratified group cross-validation. Discrimination was summarized by AUC and AP, and calibration by Brier score, calibration slope, and intercept. A prespecified count-alone comparator using the whole-lung total micronodule count was assessed on the same folds.

RESULTS

Combining fractal dimension with count yielded high discrimination between 0/1 and 1/0-1/1. For most-affected-lobe aggregation, AUC was 0.972 (95% CI, 0.927-1.000) and AP was 0.975; for whole-lung aggregation, AUC was 0.965 (95% CI, 0.921-1.000) and AP was 0.973; for the count-alone comparator, AUC was 0.917 (95% CI, 0.842-0.993) and AP was 0.947. Calibration was numerically favorable for the combined models in terms of Brier score and calibration slope.

CONCLUSIONS

Micronodule count combined with fractal dimension achieved high internal discrimination for classifying borderline CT categories. External validation and automation are needed to assess generalizability.