Although it is important to determine whether computer classification can achieve an accuracy level that is comparable to or even better than the accuracy of radiologists, it is more important to identify how computer classification might improve radiologists' interpretation of mammograms. To this end, it is clear that computer classification can potentially reduce the number of biopsies performed on benign lesions while maintaining sensitivity in the biopsy of breast cancers. Wu et al. reported that using an appropriate threshold to their computer classification results, one could hypothetically recommend biopsy in 100% of malignant cases and in only 41% of benign cases. Alternatively, using a different threshold, one could recommend biopsy in 92% of malignant cases and in only 29% of benign cases . Baker etal. reported similar findings in greater detail. For example, they showed that one could hypothetically recommend biopsy in 95% of malignant cases and in 38% of benign cases. This result was better than that could be obtained from radiologists' interpretation, which would be to recommend biopsy in 95% of cancers and 70% of benign lesions, with a statistically significant difference (p < 0.01) .
These results indicate the potential of computer classification to reduce the number of biopsies on benign lesions. However, we must remember that these results were based entirely on the computer analysis. These results do not measure how a radiologist—who makes the biopsy or follow-up recommendation — might use the computer results in making his or her own recommendation decision. We must also exercise care when interpreting the computer performance when it correctly classifies 100% of cancers in a particular database. Although it is highly desirable for computer classification to identify 100% of cancers, that extreme performance is difficult to measure accurately from a statistical point of view.
Was this article helpful?