To assess the performance of automatic lung nodule detection algorithm (research prototype) on the unseen data sets containing ground-glass opacity nodules (GGNs) and part solid nodules (PSNs).

The unseen data sets consist of 102 thin-section (0.67-1mm) MDCT scans: 101 Philips (Brilliance 64, and Mx8000 IDT 16 with the D and L reconstructions), and 1 SIEMENS (Sensation 16 with the B50f convolution kernel). Each image volume was reviewed and marked by an experienced thoracic radiologist with size and certainty (or confidence). The radiologist also defined the types of each lesion as either GGNs or PSNs. In this process, 168 nodules were found.

We selected 112 nodules (67 GGNs and 45 PSNs) out of 168 based on the criteria of confidence (>=50%) and size (6-30mm for GGNs and 3-30mm for PSNs) for the assessment of the lung nodule detection algorithm. 61 lesions (22 GGNs and 39 PSNs) out of 112 were proved as premalignant or malignant lesions including atypical adenomatous hyperplasia (AAH, n=6), bronchioloalveolar cell carcinoma (BAC, n=12), and adenocarcinoma (ADC, n=43). The other 51 lesions were not confirmed pathologically.

The lung nodule detection algorithm has been developed with a large development set of MDCT lung volumes that exclude the unseen data sets. The training sets are the mixture of SIEMENS, GE, and Philips with thin-section CT images (1-1.25mm) and are from several different scanners and different reconstruction methods such as Sensation 16, Sensation 64 (B50, B60 convolution kernels), Mx8000 IDT 16, Brilliance 16P (B reconstruction), LightSpeed Ultra, LightSpeed Pro 16. The training nodule sets consisted of 7.81% GGNs, 13.03% PSNs and 79.15% solid nodules.

The sensitivity was estimated for GGNs and PSNs. The overall sensitivity for all ground truth is 81.25% (91/112 nodules) at 5 FPs per case. The sensitivities for each category of GGNs and PSNs are 77.61% (52/67 nodules) and 86.67% (39/45 nodules) at 5 FPs, respectively.
The pathology proven nodules showed 83.61% (51/61 nodules) sensitivity at 5 FPs. The pathology proven missed lesions consist of 1 AAH, 2 BACs, and 7 ADCs.

The CAD system performance is robust on the pathology proven lesions including GGNs and PSNs. Furthermore, the system is not biased to any pathology, since the pathology ratio of the missed nodules is very similar to the ratio of the ground truth.

CAD for GGNs and PSNs