| Autor: |
Pratt, H, Coenen, F, Harding, SP, Broadbent, DM, Zheng, Y |
| Jazyk: |
angličtina |
| Rok vydání: |
2019 |
| Zdroj: |
CEUR Workshop Proceedings |
| Popis: |
Convolutional Neural Networks (CNNs) have been demonstrated to achieve state-of-the-art results on complex computer vision tasks, including medical image diagnosis of Diabetic Retinopathy (DR). CNNs are powerful because they determine relevant image features automatically. However, the current inability to demonstrate what these features are has led to CNNs being considered to be 'black box' methods whose results should not be trusted. This paper presents a method for identifying the learned features of a CNN and applies it in the context of the diagnosis of DR in fundus images using the well-known DenseNet. We train the CNN to diagnose and determine the severity of DR and then successfully extract feature maps from the CNN which identify the regions and features of the images which have led most strongly to the CNN prediction. This feature extraction process has great potential, particularly for encouraging confidence in CNN approaches from users and clinicians, and can aid in the further development of CNN methods. There is also potential for determining previously unidentified features which may contribute to a classification. |
| Databáze: |
OpenAIRE |
| Externí odkaz: |
|
