New patch-based strategy for COVID-19 automatic identification using chest x-ray images.

Purpose: The development of a robust model for automatic identification of COVID-19 based on chest x-rays has been a widely addressed topic over the last couple of years; however, the scarcity of good quality images sets, and their limited size, have proven to be an important obstacle to obtain reliable models. In fact, models proposed so far have suffered from over-fitting erroneous features instead of learning lung features, a phenomenon known as shortcut learning. In this research, a new image classification methodology is proposed that attempts to mitigate this problem. Methods: To this end, annotation by expert radiologists of a set of images was performed. The lung region was then segmented and a new classification strategy based on a patch partitioning that improves the resolution of the convolution neural network is proposed. In addition, a set of native images, used as an external evaluation set, is released. Results: The best results were obtained for the 6-patch splitting variant with 0.887 accuracy, 0.85 recall and 0.848 F1score on the external validation set. Conclusion: The results show that the proposed new strategy maintains similar values between internal and external validation, which gives our model generalization power, making it available for use in hospital settings. Supplementary Information: The online version contains supplementary material available at 10.1007/s12553-022-00704-4.

New morphological features based on the Sholl analysis for automatic classification of traced neurons.

BACKGROUND: This article addresses the automatic classification of reconstructed neurons through their morphological features. The purpose was to extend the capabilities of the L-Measure software. METHODS: New morphological features were developed, based on modifications of the conventional Sholl analysis. The lengths of the compartments, as well as their volumes, were added to the features used in the classical analysis in order to improve the results during automatic neuron classification. FSM were used to obtain subsets of lower cardinality from the full feature sets and the usefulness of these subsets was tested through their use in supervised classification tasks. The study was based on two types of neurons belonging to mice: pyramidal and GABAergic interneurons. Furthermore, a set of pyramidal neurons belonging to Later 4 and Layer 5 was analyzed. RESULTS: RF classifier shown the best performance combined with a Wrapper method.U-WNAD set allowed to obtain higher values than WN, A and D in all cases and better results than LM for the filters and wrappers FSM. U-LM-WNAD set, led to the highest AUC values for all the FSM studied. Similar results for different regions of cortex were obtained. Comparison with Existing Methods The new features exhibited high discriminatory power with which the values of AUC and Acc obtained in the experiments exceeded those obtained using only the features provided by L-Measure. CONCLUSIONS: The highest values of AUC and Acc were obtained from the sets U-WNAD and U-LM-WNAD, evidencing the discriminatory power of the new proposed features.

Feature selection for the classification of traced neurons.

BACKGROUND: The great availability of computational tools to calculate the properties of traced neurons leads to the existence of many descriptors which allow the automated classification of neurons from these reconstructions. This situation determines the necessity to eliminate irrelevant features as well as making a selection of the most appropriate among them, in order to improve the quality of the classification obtained. METHODS: The dataset used contains a total of 318 traced neurons, classified by human experts in 192 GABAergic interneurons and 126 pyramidal cells. The features were extracted by means of the L-measure software, which is one of the most used computational tools in neuroinformatics to quantify traced neurons. We review some current feature selection techniques as filter, wrapper, embedded and ensemble methods. The stability of the feature selection methods was measured. For the ensemble methods, several aggregation methods based on different metrics were applied to combine the subsets obtained during the feature selection process. RESULTS: The subsets obtained applying feature selection methods were evaluated using supervised classifiers, among which Random Forest, C4.5, SVM, Naïve Bayes, Knn, Decision Table and the Logistic classifier were used as classification algorithms. COMPARISON WITH EXISTING METHODS: Feature selection methods of types filter, embedded, wrappers and ensembles were compared and the subsets returned were tested in classification tasks for different classification algorithms. CONCLUSIONS: L-measure features EucDistanceSD, PathDistanceSD, Branch_pathlengthAve, Branch_pathlengthSD and EucDistanceAve were present in more than 60% of the selected subsets which provides evidence about their importance in the classification of this neurons.