A comparison of decision tree ensemble creation techniques.

We experimentally evaluate bagging and seven other randomization-based approaches to creating an ensemble of decision tree classifiers. Statistical tests were performed on experimental results from 57 publicly available data sets. When cross-validation comparisons were tested for statistical significance, the best method was statistically more accurate than bagging on only eight of the 57 data sets. Alternatively, examining the average ranks of the algorithms across the group of data sets, we find that boosting, random forests, and randomized trees are statistically significantly better than bagging. Because our results suggest that using an appropriate ensemble size is important, we introduce an algorithm that decides when a sufficient number of classifiers has been created for an ensemble. Our algorithm uses the out-of-bag error estimate, and is shown to result in an accurate ensemble for those methods that incorporate bagging into the construction of the ensemble.

Computer-aided mammographic screening for spiculated lesions.

PURPOSE: To study the use of a computer vision method as a second reader for the detection of spiculated lesions on screening mammograms. MATERIALS AND METHODS: An algorithmic computer process for the detection of spiculated lesions on digitized screen-film mammograms was applied to 85 four-view clinical cases: 36 cases with cancer proved by means of biopsy and 49 cases with negative findings at examination and follow-up. The computer detections were printed as film with added outlines that indicated the suspected cancers. Four radiologists screened the 85 cases twice, once without and once with the computer reports as ancillary films. RESULTS: The algorithm alone achieved 100% sensitivity, with a specificity of 82%. The computer reports increased the average radiologist sensitivity by 9.7% (P = .005), moving from 80.6% to 90.3%, with no decrease in average specificity. CONCLUSION: The study demonstrated that computer analysis of mammograms can provide a substantial and statistically significant increase in radiologist screening efficacy.

Advances in the automated detection of metaphase chromosomes labeled with fluorescence dyes.

Applications of fluorescence in situ hybridization (FISH) for translocation studies and biological dosimetry would benefit substantially from reliable and efficient automatic detection of metaphase chromosomes labeled with fluorescent dyes. We replicated and evaluated a fluorescence metaphase finder previously developed at the Medical Research Council (MRC), Human Genetics Unit (Scotland) and at Lawrence Berkeley Laboratory (LBL; California). The MRC/LBL system seemed to detect nearly all of the metaphases on the test slides, but it presented an unacceptable number of false positives (about five false positives per one true positive). Furthermore, we determined that the system actually overcalled true detections by counting certain metaphase spreads twice (duplicates). Through modifications of the MRC/LBL system, we developed the Lawrence Livermore National Laboratory (LLNL) system, which minimizes the detection of duplicates, incorporates new detection features, uses a binary decision tree (BDT) for classification, and provides functionalities to improve scanning accuracy and improve the post-detection review. To test the new system, DAPI-stained preparations of metaphase chromosomes from blood lymphocytes of four unrelated donors were placed on slides in drops ranging from 7 mm to 20 mm in diameter. Drops contained between 5 and 200 scorable metaphases each. The LLNL system achieved approximately 90% detection of non-duplicated metaphases as verified by an expert cytogeneticist, with typically less than one false positive per every one true positive detected.