Statistical Learning without Ground Truth in Computer Vision and Medicine

Опубликовано 17 августа 2016, 21:16

The first part of the talk will present and overview of our machine learning endeavors within the scope of Computational Pathology: (i) generating a gold standard based on clinical labeling experiments, (ii) using off-line and on-line ensemble learning techniques to train object detectors for cell nuclei and (iii) employing Bayesian survival statistics for biomarker detection and diagnosing cancer patients. Based on several interdisciplinary research projects I will demonstrate how insights gained from statistical modeling can be translated to biomedical knowledge and in which ways clinical decision making can benefit from it. In the second part of the talk I am going to give an outlook on learning under labeling uncertainty: In a large number of real world application an objective ground truth is not available or too expensive to acquire. In practice, as a last resort, one would ask several domain experts for their opinion about each object in question to generate a gold standard. Depending on the difficulty of the task this often results in ambiguous labeling due to disagreement between experts. The resulting labeling matrix poses a non-trivial challenge for supervised learning. We will investigate under which condition it is possible to learn more about the data generating distribution than just using majority vote. A positive result would have immense influence in domains where specific models can be trained for years by a large number of experts, e.g. medical decision support. I will illustrate the problem with examples from medicine and space exploration, i.e. the classification of cell nuclei and the recognition of volcanoes on Venus.