CAD Performance Analysis for Pulmonary Nodule Detection: Comparison of Thick- and Thin-Slice Helical CT Scans Randy D Ernst 1, Russell C Hardie 2, Metin N Gurcan 3, Aytekin Oto 1, Steve K Rogers 3, Jeffrey W Hoffmeister 3 1. Department of Radiology, The University of Texas Medical Branch, Galveston TX 2. iCAD Inc. and University of Dayton, Dayton OH 3. iCAD Inc., Beavercreek OH
Introduction This study compares the performance of a CAD (QuickCue™, iCAD, Inc.) system in detecting lung nodules from thick- and thin-slice multi-detector row CT scans, and to evaluate the potential benefit of CAD on radiologist sensitivity.
Methods and Materials 57 reports reviewed retrospectively Case selection: Obtained during a 5-month period Referred from multiple departments Contain at least 1 pulmonary nodule but fewer than 10 nodules to localize Have no significant breathing miss - registration, post surgical changes, pleural effusions & atelectasis
Methods and Materials 4-detector LightSpeed QX/I Scanner, GE systems HQ setting with 5.0 collimation, helical pitch of 0.75/1.0 Standard-dose ( mA, 120 kVp) Images were reconstructed at 5 mm (thick) and 2.5 mm (thin) slice thicknesses.
Methods and Materials 140 nodules (3 mm - 25 mm) were identified pre-CAD by radiologists From thick-slice cases only. Cases with multiple nodules were excluded. Truth marks were mapped to the thin-slice data Mean nodule size 7.3 ± 4.2 mm (3 – 25 mm) Gold standard for nodule truth comes for post- CAD Radiologist review One gold standard for thick-slice and one for thin- slice
CAD System (QuickCue™, iCAD Inc.) 3D Lung Segmentation 3D Candidate Segmentation Calculate Features DICOM Images Classifier Detection Mask
CAD detected 72.1% (101/140) of the pre-CAD truth nodules CAD detected 35 additional radiologist- confirmed nodules, an increase of 25% (35/140) in sensitivity 5.6 (317/57) false-positives per case 55 due to atelectasis 18 due to scarring Review of Thick-Slice CAD Results
Venn Diagram for Thick CAD Pre-CAD Review Post-CAD Review Gold Standard
CAD detected 80.7% (113/140) of the pre-CAD truth nodules. CAD detected 94 additional radiologist- confirmed nodules, an increase of 67.1% (94/140). 4.6 (262/57) false-positives per case. 70 due to atelectasis 39 due to scarring Review of Thin-Slice CAD Results
Venn Diagram for Thin CAD using thin-slice Pre-CAD Review using thick-slice with detections mapped to thin-slice Post-CAD Review of thin-slice Gold Standard
Comparison Thick-slice cases Thin-slice cases CAD sensitivity 72.1%80.7% Radiologist sensitivity increase after CAD +25%+67.1% FPs5.64.6
FROC Curve for CAD
CAD detections in Thick-Slice Additional Detections
CAD detections in Thin-Slice Additional Detections
5 primary lung cancers 24 cases of metastatic cancer including 7 lymphomas, 4 breast, 4 head and neck, 2 colon, 2 pancreas, 1 carcinoid, 1 seminoma, 1 ovarian, 1 melanoma and 1 tracheal papillomatosis 23 cases of infection, including 19 granulomatous disease either calcified, stable on follow-up or biopsy proven. 4 were presumed infection that resolved with follow- up 1 case proved to be a thrombosed AVM 4 cases lost to follow up Case Follow-up
Example TPs Examples of nodules that are detected by both radiologist and CAD
Example TPs Examples of nodules that are initially missed by radiologists then detected after reviewing CAD
Review of CAD Results Sources of false positives Vessel intersections Inaccurate lung segmentation Partial volume effects Other lung abnormalities (scarring, atelectasis)
Example FPs
Review of CAD Results Sources of false negatives (missed nodules) Low density, irregular Strong connectivity with vessels Imperfect candidate segmentation Inaccurate lung segmentation
Example FNs
Conclusions Preliminary results indicate that both sensitivity and specificity of the CAD system increases when used with thin- slice scans versus thick-slice scans. The CAD system operating on both thick- and thin-slice scans improved radiologist sensitivity Improvement was greater for CAD operating on thin-slice scans