Japanese researchers have developed a new a computer-aided diagnosis (CADx) system that’s superior to some other methods, relatively easy to use and capable of differentiating between malignant and benign nodules on lung CT.

Academic Radiology published their report online Jan. 16.

Mizuho Nishio, MD, PhD, and Chihiro Nagashima, MD, both of the Institute of Biomedical Research and Innovation in Hyogo Prefecture, analyzed 73 lung nodules revealed on 60 sets of CT images, 46 of which were from contrast-enhanced scans.

The images were provided by the LUNGx Challenge. As the ground truth of the lung nodules was unavailable, the researchers used radiological evaluation to construct a surrogate ground truth.

Their experimental CADx technique combined novel patch-based feature extraction with principal component analysis, image convolution and pooling operations.

Nishio and Nagashima compared the performance of their system against that of three established systems: histogram of CT density, local binary pattern on three orthogonal planes and 3D random local binary pattern.

They analyzed the probabilistic outputs of the systems and surrogate ground using receiver operating characteristic analysis and area under the curve (AUC) calculation.

In addition, a LUNGx Challenge team calculated the AUC of Nishio and Nagashima’s proposed method based on the actual ground truth of the LUNGx Challenge team’s dataset.

The tested technique yielded a better AUC than the comparison methods, and it proved comparable to that calculated by the LUNGx Challenge team using the actual ground truth:

• Based on the surrogate ground truth, the areas under the curve were as follows: histogram of CT density, 0.640; local binary pattern on three orthogonal planes, 0.688; three-dimensional random local binary pattern, 0.725; and Nishio and Nagashima’s proposed method, 0.837.
• Based on the actual ground truth, the area under the curve of the new method was 0.81.

In their discussion, Nishio and Nagashima state that the main advantage of their proposed method is its simplicity, as it precludes the need to segment a lung nodule or input radiological or clinical findings into the CADx system.

They cite previous studies in which successful differentiation of lung nodules required nodule segmentation, calculation of many types of image features or radiological findings.

By contrast, only one type of image feature was used in Nishio and Nagashima’s proposed method—and, they underscore, this did not require nodule segmentation.

“While it was easy to implement the proposed method, this method produced the best diagnostic accuracy among the four methods evaluated, obtaining discriminative characteristics of lung nodules,” they write.

“Although it is difficult to compare our results directly to those of other CADx systems, the proposed method was comparable to state-of-the-art CADx systems on the basis of the results of these studies,” the authors add. “Although the results of the present study are preliminary, we expect that our CADx system will be useful for clinical use.”