Linn Fredriksson och Sara Lundell, Elektroteknik

​Titel: Weak Supervision for Medical Image Analysis: Image Classification Compared in Supervised and Weakly Supervised Learning Settings
Opponenter: Alma Lund och Felix Ericsson
Examinator: Ida Häggström

Abstract:
For medical disease diagnosis, studies have shown that artificial intelligence-based models can be trained to perform just as effectively or even outperform humans. Such models are typically trained in a supervised manner that requires large amounts of data with ground truth labels. Hence, a major bottleneck is the laborious labeling/annotation that is required. A proposed way to overcome this bottleneck is weak supervision. Weak supervision uses limited or noisy data as training data in a supervised learning setting.

In this project, the performance of an image classification model trained in both a classically supervised setting and a weakly supervised setting was measured and compared. A convolutional neural network model was used as an image classifier. Images of skin lesions were classified as malignant or benign melanoma. The noisy data was programmatically annotated by using heuristics and rules. So-called labeling functions were used to implement these rules which together with statistical models generate probabilistic labels for the images. The labeling functions were formulated for metadata, such as age, gender and character traits of the mole. 2000 images were used as training and validation data and 600 images were set aside for test data. The weakly annotated data receives an averaged area under the curve score at 0.82, and for the ground truth data, 0.86 on the test data. The weakly supervised model shows promising results, scoring an accuracy of 76%. The result shows that weak supervision can create good predictive image classifiers. Hence, these models have the potential to be used to generate new training data that can be used for a image classifier.

Welcome!
Linn, Sara and Ida
Password: 317919
Kategori Studentarbete
Plats: EDIT-rummet 3364 (EDIT-huset)
Tid: 2022-06-02 09:00
Sluttid: 2022-06-02 10:00

Sidansvarig Publicerad: fr 20 maj 2022.