guest ::
| Home > Publications Database > Abstract: Leveraging Web Data for Skin Lesion Classification |
| Home > Publications Database > Abstract: Leveraging Web Data for Skin Lesion Classification |
| Contribution to a conference proceedings/Contribution to a book | DZNE-2022-01056 |
; ; ; ; ; ; ; ; ;
2019
Springer Fachmedien Wiesbaden
Wiesbaden
ISBN: 978-3-658-25325-7 (print), 978-3-658-25326-4 (electronic)
This record in other databases:
Please use a persistent id in citations: doi:10.1007/978-3-658-25326-4_44
Abstract: The success of deep learning is mainly based on the assumption that forthe given application, there is access to a large amount of annotated data. Inmedical imaging applications, having access to a big-well-annotated data-set isrestrictive, time-consuming and costly to obtain. Although diverse techniques asdata augmentation can be leveraged to increase the size and variability within thedata-set, the representativeness of the training set is still limited by the numberof available samples. Furthermore, a small-size and well-annotated data-set cannot guarantee the generalizability to unseen samples.As a consequence for the aforementioned problem, we have proposed in [1]to utilize the vast amount of free available data from the web to alleviate theneed of a large-well-annotated data-set in the so-called Webly Supervised Learning methodology presented in [2]. Harvesting images from the web presents theopportunity to increase the variability and heterogeneity of the training set atthe cost of label noise. These label noises include cross-domain: retrieved imagesopposite to the dermatology domain and cross-category: retrieved images visually similar to the query image yet belonging to a different class. To overcomethe first type of noise we have proposed a search by image technique to increasethe search specificity and retrieve only images visually similar to the query. Thesecond type of noise is reduced by modeling the noise in the data-set with a classtransition matrix, estimated from the web-retrieved images as proposed in [3].To the best of our knowledge, our work has been the first applying webly supervised learning in medical imaging. To validate our methodology, we have testedour system in the context of ten-class fine-grained skin lesion classification. Ourresults show that the proposed methodology increase the overall classificationaccuracy from 71.25 % to 80.53 % due to the web-supervision.
|
The record appears in these collections: |