Effect of CT Acquisition Parameters on RetinaNet-Based Lung Nodule Detection: A Physics-Guided Validation Study Across Dose, Slice Thickness, and Reconstruction Kernel

Abstract

Rationale and Objectives

Computed tomography acquisition parameters, including dose level, slice thickness, and reconstruction kernel, vary substantially across clinical sites, yet artificial intelligence (AI)-assisted lung nodule detection systems are rarely evaluated against this variability. We propose and demonstrate a physics-guided framework for evaluating detection model sensitivity to systematic acquisition parameter variation.

Computed tomography acquisition parameters, including dose level, slice thickness, and reconstruction kernel, vary substantially across clinical sites, yet artificial intelligence (AI)-assisted lung nodule detection systems are rarely evaluated against this variability. We propose and demonstrate a physics-guided framework for evaluating detection model sensitivity to systematic acquisition parameter variation.

Materials and Methods

154 cases from the LIDC-IDRI dataset were evaluated using a Medical Open Network for AI (MONAI) RetinaNet model pretrained on LUNA16 (fold 0, no fine-tuning) across six imaging conditions: baseline, 25% dose reduction, 50% dose reduction, 3 mm slice thickness, 5 mm slice thickness, and soft kernel reconstruction. Dose reduction was simulated via image-domain Gaussian noise; slice thickness via z-axis moving average. Detection sensitivity was computed at confidence threshold 0.5 with a 15 mm matching criterion.

154 cases from the LIDC-IDRI dataset were evaluated using a Medical Open Network for AI (MONAI) RetinaNet model pretrained on LUNA16 (fold 0, no fine-tuning) across six imaging conditions: baseline, 25% dose reduction, 50% dose reduction, 3 mm slice thickness, 5 mm slice thickness, and soft kernel reconstruction. Dose reduction was simulated via image-domain Gaussian noise; slice thickness via z-axis moving average. Detection sensitivity was computed at confidence threshold 0.5 with a 15 mm matching criterion.

Results

Baseline sensitivity was 84.8% (95% CI: 80.2–89.8%). Soft kernel and 5 mm slice thickness produced the largest decreases: 74.3% (−10.5 pp, p < 0.05) and 71.6% (−13.2 pp, p < 0.05), respectively. Dose reduction caused moderate asymmetric degradation: 76.0% at 25% dose (−8.8 pp, p < 0.05) and 80.0% at 50% dose (−4.8 pp, p < 0.05). 3 mm slice thickness produced a small but significant reduction (−2.4 pp, p = 0.009). Reductions were most pronounced in the 3–6 mm range and consistent across confidence thresholds.

Baseline sensitivity was 84.8% (95% CI: 80.2–89.8%). Soft kernel and 5 mm slice thickness produced the largest decreases: 74.3% (−10.5 pp, p < 0.05) and 71.6% (−13.2 pp, p < 0.05), respectively. Dose reduction caused moderate asymmetric degradation: 76.0% at 25% dose (−8.8 pp, p < 0.05) and 80.0% at 50% dose (−4.8 pp, p < 0.05). 3 mm slice thickness produced a small but significant reduction (−2.4 pp, p = 0.009). Reductions were most pronounced in the 3–6 mm range and consistent across confidence thresholds.

Conclusion

Slice thickness and soft reconstruction kernel represent stronger constraints on AI detection performance than image noise under these conditions. The proposed framework is reproducible, requires no proprietary scanner data, and provides a practical basis for post-deployment acquisition QA.

Slice thickness and soft reconstruction kernel represent stronger constraints on AI detection performance than image noise under these conditions. The proposed framework is reproducible, requires no proprietary scanner data, and provides a practical basis for post-deployment acquisition QA.