With the increasing advancements in artificial intelligence and the growing collections of medical data, new methods are emerging that could help doctors with diagnosis. This could help improve the detection of diseases and decrease the workload of healthcare professionals. Despite the success of some models in classifying medical data, the issue of transparency in these models remains. Deep neural network models contain a large number of parameters and are difficult to interpret. In the field of medicine, incorrect decisions can have long-lasting and severe consequences, making trust in the model's predictions even more critical than in other fields. Although some models achieve excellent results, we cannot blindly trust them without first understanding why the model decided for a given diagnosis. In this master's thesis, we focused on the area of mammography images and trained models to diagnose images as healthy or cancerous. The models achieved AUC values under the ROC curve above 0.90. We then examined the results and explanations of different models on several test mammography images. During the review of the results, we observed differences in the explanations provided by various interpretability methods for the same models, as well as for different architectures, despite the same diagnosis. The methods used for explanations in this study were found to be ineffective in the field of mammography.
|
