Abstrakt
We are developing a complex computer aided diagnosis (CAD) system to detect small pulmonary nodules from helical CT scans. Here we present a classifier to reduce the number of false positive responses of the primary detector.
Our approach is based on an asymmetric Adaboost which enables us to give different weights to missed nodules (false negatives, FNs) and incorrectly detected structures (false positives, FPs). This is useful because there are noticeably more negative examples in the nodule candidate set than real nodules-true positives (TPs).
The whole system is meant as a second opinion for a human radiologist to speed up reading the examination. That is why we should detect as many true nodules as possible, while a certain number of FPs is acceptable.
The system was tested on 147 cases (36559 slices) containing 357 nodules marked by an expert radiologist. The new classifier significantly reduced the number of false positives, while only a few nodules were incorrectly omitted.