Sparse multi-label classification of medical images using deep convolutional neural networks
The recent advancements in imaging technologies have improved clinician's ability to detect, diagnose, and treat diseases. As an example, radiologists routinely interpret medical images and summarize their findings in the form of radiology reports. The mapping of visual information present in m...
Αποθηκεύτηκε σε:
| Κύριοι συγγραφείς: | , |
|---|---|
| Άλλοι συγγραφείς: | |
| Γλώσσα: | English |
| Δημοσίευση: |
2021
|
| Θέματα: | |
| Διαθέσιμο Online: | http://hdl.handle.net/11610/21660 |
| Ετικέτες: |
Προσθήκη ετικέτας
Δεν υπάρχουν, Καταχωρήστε ετικέτα πρώτοι!
|
| Περίληψη: | The recent advancements in imaging technologies have improved clinician's ability to detect, diagnose, and treat diseases. As an example, radiologists routinely interpret medical images and summarize their findings in the form of radiology reports. The mapping of visual information present in medical images to the condensed textual description is a tedious, time-consuming, expensive, and error-prone task. The development of methods that can automatically detect the presence and location of medical concepts in medical images can improve radiologists' efficiency, reduce the burden of manual interpretation, and help reduce diagnostic errors.
In this master thesis, we deal with the challenging task of medical image tagging (or labeling), which aims to identify medical terms (or concepts) in medical images, and with the ultimate goal of helping physicians to generate medical reports from medical images. In particular, we propose a variation of convolutional neural networks for sparse multi-label classification to predict relevant concepts present in images, and we test it against all recent datasets from the ImageCLEFmed Caption task (i.e., 2017, 2018, 2019, and 2020). The proposed system outperformed all winning teams in terms of F1 score between system predicted and ground truth concepts. We present our work with data analysis, experimental results, comparisons between the different hyper-parameters and network architectures, and last but not least, with a short discussion on future steps. |
|---|