Keywords: Lesion Segmentation, Weakly Supervised Learning, Multi-Task Learning
TL;DR: Multi-task learning helps weakly supervised CT and skin lesion segmentation
Abstract: We introduce the concept of multi-task learning to weakly-supervised lesion segmentation, one of the most critical and challenging tasks in medical imaging. Due to the lesions' heterogeneous nature, it is difficult for machine learning models to capture the corresponding variability. We propose to jointly train a lesion segmentation model and a lesion classifier in a multi-task learning fashion, where the supervision of the latter is obtained by clustering the RECIST measurements of the lesions. We evaluate our approach specifically on liver lesion segmentation and more generally on lesion segmentation in computed tomography (CT), as well as segmentation of skin lesions from dermatoscopic images. We show that the proposed joint training improves the quality of the lesion segmentation by 4% percent according to the Dice coefficient and 6% according to averaged Hausdorff distance (AVD), while reducing the training time required by up to 75%.
Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.
Source Code Url: https://github.com/esizikova/weaklesionmtl
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.
Data Set Url: https://pubmed.ncbi.nlm.nih.gov/30035154/ https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/DBW86T https://arxiv.org/abs/1901.04056
Paper Type: methodological development
Source Latex: zip
Primary Subject Area: Application: Radiology
Secondary Subject Area: Learning with Noisy Labels and Limited Data