Ensemble of convolutional neural networks for classification of breast microcalcification from mammograms.

Human level recall performance in detecting breast cancer considering microcalcifications from mammograms has a recall value between 74.5% and 92.3%. In this research, we approach to breast microcalcification classification problem using convolutional neural networks along with various preprocessing methods such as contrast scaling, dilation, cropping etc. and decision fusion using ensemble of networks. Various experiments on Digital Database for Screening Mammography dataset showed that preprocessing poses great importance on the classification performance. The stand-alone models using the dilation and cropping preprocessing techniques achieved the highest recall value of 91.3%. The ensembles of the stand-alone models surpass this recall value and a 97.3% value of recall is achieved. The ensemble having the highest F1 Score (harmonic mean of precision and recall), which is 94.5%, has a recall value of 94.0% and a precision value of 95.0%. This recall is still above human level performance and the models achieve competitive results in terms of accuracy, precision, recall and F1 score measures.

A Bayesian Approach to Learning Scoring Systems.

We present a Bayesian method for building scoring systems, which are linear models with coefficients that have very few significant digits. Usually the construction of scoring systems involve manual effort-humans invent the full scoring system without using data, or they choose how logistic regression coefficients should be scaled and rounded to produce a scoring system. These kinds of heuristics lead to suboptimal solutions. Our approach is different in that humans need only specify the prior over what the coefficients should look like, and the scoring system is learned from data. For this approach, we provide a Metropolis-Hastings sampler that tends to pull the coefficient values toward their "natural scale." Empirically, the proposed method achieves a high degree of interpretability of the models while maintaining competitive generalization performances.

Nonconvex online support vector machines.

In this paper, we propose a nonconvex online Support Vector Machine (SVM) algorithm (LASVM-NC) based on the Ramp Loss, which has the strong ability of suppressing the influence of outliers. Then, again in the online learning setting, we propose an outlier filtering mechanism (LASVM-I) based on approximating nonconvex behavior in convex optimization. These two algorithms are built upon another novel SVM algorithm (LASVM-G) that is capable of generating accurate intermediate models in its iterative steps by leveraging the duality gap. We present experimental results that demonstrate the merit of our frameworks in achieving significant robustness to outliers in noisy data classification where mislabeled training instances are in abundance. Experimental evaluation shows that the proposed approaches yield a more scalable online SVM algorithm with sparser models and less computational running time, both in the training and recognition phases, without sacrificing generalization performance.We also point out the relation between nonconvex optimization and min-margin active learning.