Skin lesions are caused due to multiple factors, like allergies, infections, exposition to the sun, etc. These skin diseases have become a challenge in medical diagnosis due to visual similarities, where image classification is an essential task to achieve an adequate diagnostic of different lesions. Melanoma is one of the best-known types of skin lesions due to the vast majority of skin cancer deaths. In this work, we propose an ensemble of improved convolutional neural networks combined with a test-time regularly spaced shifting technique for skin lesion classification. The shifting technique builds several versions of the test input image, which are shifted by displacement vectors that lie on a regular lattice in the plane of possible shifts. These shifted versions of the test image are subsequently passed on to each of the classifiers of an ensemble. Finally, all the outputs from the classifiers are combined to yield the final result. Experiment results show a significant improvement on the well-known HAM10000 dataset in terms of accuracy and F-score. In particular, it is demonstrated that our combination of ensembles with test-time regularly spaced shifting yields better performance than any of the two methods when applied alone.

In this work we implement a COVID-19 infection detection system based on chest X-ray images with uncertainty estimation. Uncertainty estimation is vital for safe usage of computer aided diagnosis tools in medical applications. Model estimations with high uncertainty should be carefully analyzed by a trained radiologist. We aim to improve uncertainty estimations using unlabelled data through the MixMatch semi-supervised framework. We test popular uncertainty estimation approaches, comprising Softmax scores, Monte-Carlo dropout and deterministic uncertainty quantification. To compare the reliability of the uncertainty estimates, we propose the usage of the Jensen-Shannon distance between the uncertainty distributions of correct and incorrect estimations. This metric is statistically relevant, unlike most previously used metrics, which often ignore the distribution of the uncertainty estimations. Our test results show a significant improvement in uncertainty estimates when using unlabelled data. The best results are obtained with the use of the Monte Carlo dropout method.

The application of deep learning to image and video processing has become increasingly popular nowadays. Employing well-known pre-trained neural networks for detecting and classifying objects in images is beneficial in a wide range of application fields. However, diverse impediments may degrade the performance achieved by those neural networks. Particularly, Gaussian noise and brightness, among others, may be presented on images as sensor noise due to the limitations of image acquisition devices. In this work, we study the effect of the most representative noise types and brightness alterations on images in the performance of several state-of-the-art object detectors, such as YOLO or Faster-RCNN. Different experiments have been carried out and the results demonstrate how these adversities deteriorate their performance. Moreover, it is found that the size of objects to be detected is a factor that, together with noise and brightness factors, has a considerable impact on their performance.