Optimal reconstruction and quantitative image features for computer-aided diagnosis tools for breast CT

Area of Science:

• Medical Imaging
• Radiology
• Computer-Aided Diagnosis

Background:

• Dedicated breast computed tomography (bCT) is an emerging imaging modality.
• Developing effective computer-aided diagnosis (CADx) schemes is crucial for improving diagnostic accuracy in bCT.
• Image reconstruction quality significantly impacts the performance of CADx systems.

Purpose of the Study:

• To identify the optimal image reconstruction and quantitative feature set for a CADx system in bCT.
• To evaluate the diagnostic performance of the optimized CADx system against radiologists.

Main Methods:

• Utilized 93 bCT scans with 102 breast lesions (62 malignant, 40 benign).
• Generated 38 image reconstructions per case using an iterative image reconstruction (IIR) algorithm, varying image sharpness.
• Extracted 23 quantitative image features post-lesion segmentation and employed leave-one-out cross-validation (LOOCV) for feature selection and classifier training (linear discriminant analysis).
• Conducted an observer study comparing the CADx system's performance to six radiologists on 50 lesions.

Main Results:

• Classifier performance improved with increased image sharpness.
• An optimal sharp reconstruction and three quantitative features yielded the highest diagnostic performance in LOOCV.
• The CADx system achieved an area under the ROC curve (AUC) of 0.94, significantly outperforming radiologists (max AUC 0.78).
• The classifier's partial AUC (pAUC) at >=90% sensitivity was statistically superior to radiologists' (p<0.0001).

Conclusions:

• Image sharpness is a reliable indicator for estimating CADx diagnostic performance in bCT.
• An optimal combination of sharp reconstruction and specific feature set maximizes CADx algorithm performance.
• The developed CADx algorithm, when optimized, surpasses radiologist performance in breast lesion diagnosis using bCT.