Abstract

Abstract

Abstract
Machine learning algorithms have evolved by exchanging simplicity and interpretability for accuracy, which prevents their adoption in critical tasks such as healthcare. Progress can be made by improving interpretability of complex models while preserving performance. This work introduces an extension of interpretable mimic learning which teaches in-terpretable models to mimic predictions of complex deep neural networks, not only on binary problems but also in ordinal settings. The results show that the mimic models have comparative performance to Deep Neural Network models, with the advantage of being interpretable.

Abstract

Abstract
Although cross-validation is a standard procedure for performance evaluation, its joint application with oversampling remains an open question for researchers farther from the imbalanced data topic. A frequent experimental flaw is the application of oversampling algorithms to the entire dataset, resulting in biased models and overly-optimistic estimates. We emphasize and distinguish overoptimism from overfitting, showing that the former is associated with the cross-validation procedure, while the latter is influenced by the chosen oversampling algorithm. Furthermore, we perform a thorough empirical comparison of well-established oversampling algorithms, supported by a data complexity analysis. The best oversampling techniques seem to possess three key characteristics: use of cleaning procedures, cluster-based example synthetization and adaptive weighting of minority examples, where Synthetic Minority Oversampling Technique coupled with Tomek Links and Majority Weighted Minority Oversampling Technique stand out, being capable of increasing the discriminative power of data.

Abstract
One of the main difficulties in the use of deep learning strategies in medical contexts is the training set size. While these methods need large annotated training sets, these datasets are costly to obtain in medical contexts and suffer from intra and inter-subject variability. In the present work, two new pre-processing techniques are introduced to improve a deep classifier performance. First, data augmentation based on co-registration is suggested. Then, multi-scale enhancement based on Difference of Gaussians is proposed. Results are accessed in a public mammogram database, the InBreast, in the context of an ordinal problem, the BI-RADS classification. Moreover, a pre-trained Convolutional Neural Network with the AlexNet architecture was used as a base classifier. The multi-class classification experiments show that the proposed pipeline with the Difference of Gaussians and the data augmentation technique outperforms using the original dataset only and using the original dataset augmented by mirroring the images.