• 2022-05
  • 2022-04
  • 2020-08
  • 2020-07
  • 2018-07
  • Loxapine Succinate br The other challenging problems of breast


    The other challenging problems of breast cancer histopathology image analyses are related to the limited number of available training samples and the imbalanced data problem. First, DCNNs are only effective when the number of available training samples is large enough during a training stage. Conversely, these networks often suffer from overfitting if the train-ing samples are limited. Unlike the case of natural image classification tasks, there are far fewer medical images available for use for the effective training of a deep learning network for medical applications, such as breast cancer detection. This is be-cause of data privacy issues and the increased cost of data collection. Second, like in many other medical image applications, breast cancer detection methods are also associated with the problem of unbalanced training images, because collecting data from patients is not an easy task. To address these problems, we propose a novel boosting method for improving classifi-cation performance and for preventing overfitting. This boosting method is based on the combination of a DCNN model and a boosting trees classifier. In this combination, the DCNN model can be used to map high-dimensional inputs to a low-dimensional space of discriminative feature vectors. In fact, the classification performance of each DCNN degrades after a certain number of training iterations owing to the overfitting problem. In this situation, we can stop training the DCNN r> model early on, and use it Loxapine Succinate to extract low-dimensional discriminated features. In particular, the last convolutional layer of the DCNN model is converted into a one-dimensional feature vector with a length of 1056. Finally, this one-dimensional fea-ture vector is used as the input to the gradient boosting trees classifier. Unlike DCNNs, the boosting trees classifier is able to tackle the challenging problems of the limited number of available training samples and the imbalanced data problem. The boosting trees classifier can further improve the DCNN performance provided that the training data belongs to a low dimensional subspace. This is because the boosting trees classifier is able to convert weak classifiers to strong classifiers by adding more weight to training examples that were misclassified by weaker classifiers in earlier rounds. Furthermore, a data augmentation method is used to tackle imbalanced data, and to increase the available data samples for training the DCNNs and for boosting trees classifiers. For these reasons, the combination of DCNN and a boosting trees classifier leads to a better classification performance even though the number of breast cancer samples is not large, owing to privacy policy constraints and other conditions.
    In general, the training process for this method consists of two stages. In the first stage, we aimed at building a set of deep convolutional neural networks trained to extract the most useful visual features from multi-scale training images. Recent research demonstrated that an ensemble of DCNNs can perform better than a single DCNN [33,34]. Hence, we trained independently multiple DCNNs with different input scales and combined their final feature vectors by using a new boosting strategy. In fact, these advance feature vectors were used as inputs to gradient boosting trees classifiers [14]. In the second training stage, each gradient boosting trees classifier was trained based on the new dataset of deep feature vectors extracted from the corresponding DCNN in the first stage. After the training processes, each boosting trees classifier can achieve a better accuracy rate than its corresponding DCNN classifier. Finally, we applied the majority voting strategy for combining the boosting trees classifiers. This combination resulted in a final boosting classifier that achieved the best classification performance thanks to the improved extraction of multi-scale deep features of breast cancer tumors.
    In this study, we tested our proposed algorithms and its competitors to evaluate the accuracy of breast detection in a challenging dataset. Based on extensive experiments, our algorithms are shown to significantly outperform state-of-the-art algorithms. In particular, our proposed method provides several contributions:
    • First, we present an ensemble of DCNNs trained to extract the most useful visual features from multi-scale training-images. By building the ensemble of DCNNs, both global and local features of breast cancer tumors can be extracted properly, and the accuracy of multi-label breast cancer detection is thus significantly improved.
    • Second, we propose to use the gradient boosting trees algorithm to boost the classification performance of the DCNN classifiers. We proved that the combination of DCNN and a boosting trees classifier leads to a better classification per-formance despite the limited number of breast cancer samples and imbalanced training data.