PROJECT TITLE :

On Combining Multiple-Instance Learning and Active Learning for Computer-Aided Detection of Tuberculosis

ABSTRACT:

The major advantage of multiple-instance learning (MIL) applied to a computer-aided detection (CAD) system is that it permits optimizing the latter with case-level labels instead of accurate lesion outlines as historically required for a supervised approach. As shown in previous work, a MIL-based mostly CAD system will perform comparably to its supervised counterpart considering complex tasks such as chest radiograph scoring in tuberculosis (TB) detection. However, despite this exceptional achievement, the uncertainty inherent to MIL will cause a less satisfactory outcome if analysis at lower levels (e.g., regions or pixels) is needed. This issue could seriously compromise the applicability of MIL to tasks related to quantification or grading, or detection of highly localized lesions. During this paper, we propose to scale back uncertainty by embedding a MIL classifier at intervals a lively learning (AL) framework. To attenuate the labeling effort, we develop a completely unique instance selection mechanism that exploits the MIL problem definition through one-category classification. We have a tendency to adapt this mechanism to supply meaningful regions rather than individual instances for knowledgeable labeling, which could be a a lot of applicable strategy given the appliance domain. Additionally, and contrary to usual AL strategies, one iteration is performed. To point out the effectiveness of our approach, we compare the output of a MIL-based CAD system trained with and while not the proposed AL framework. The task is to detect textural abnormalities related to TB. Each quantitative and qualitative evaluations at the pixel level are allotted. Our technique significantly improves the MIL-primarily based classification.