Multi-task Learning in the Computerized Diagnosis of Breast Cancer on DCE-MRIs

TL;DR

This paper explores multi-task learning to improve breast cancer diagnosis from DCE-MRI images, addressing scanner heterogeneity to enhance predictive accuracy.

Contribution

It introduces a multi-task learning approach that models relationships across datasets from different MRI scanner strengths, improving classification performance.

Findings

MTL outperforms SVM on merged datasets

Higher predictive power with MTL

Addresses scanner heterogeneity in MRI data

Abstract

Hand-crafted features extracted from dynamic contrast-enhanced magnetic resonance images (DCE-MRIs) have shown strong predictive abilities in characterization of breast lesions. However, heterogeneity across medical image datasets hinders the generalizability of these features. One of the sources of the heterogeneity is the variation of MR scanner magnet strength, which has a strong influence on image quality, leading to variations in the extracted image features. Thus, statistical decision algorithms need to account for such data heterogeneity. Despite the variations, we hypothesize that there exist underlying relationships between the features extracted from the datasets acquired with different magnet strength MR scanners. We compared the use of a multi-task learning (MTL) method that incorporates those relationships during the classifier training to support vector machines run on a merged dataset that includes cases with various MRI strength images. As a result, higher predictive power is achieved with the MTL method.