Exploring the Relationship Between Error and Interpretation of the Segmentation Model's Prediction
Nikita Detkov, Ksenia Balabaeva, Sergey Valerevich Kovalchuk · Procedia Computer Science · 2022
The rapid adoption of artificial intelligence methods in the medical industry opens up more and more questions for researchers about the control of their predictions and the explanations for those predictions. One of such medical tasks is the detection of functional tissue units in microscope kidney images. In this paper we study the XAI (eXplainable Artificial Intelligence) methods related to the segmentation task, explore the techniques of analysing the model's prediction and analyse the obtained errors. During the work we studied the hypothesis of filtering errors from model's prediction using insights from it's interpretation and came up with the method of improving the dataset labeling reliability.