Name: | Description: | Size: | Format: | |
---|---|---|---|---|
8.75 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Breast cancer is a significant global health concern, affecting thousands of individuals,
primarily women, with estimated cases expected to climb by 2040. Early-stage diagnosis
is essential for effective treatment and better patient outcomes. This dissertation
presents a mammogram retrieval system based on the aggregation of image classifiers
to aid specialists in diagnosing breast cancer. The system uses a retrieval model that
combines the output of multiple classifiers, each targeting different dimensions related
to breast cancer diagnosis. These dimensions include breast density, asymmetries, BIRADS
classification, calcifications, distortions, laterality, masses, and image incidence.
This dissertation also describes the creation of an application to collect ground truth
data to aid engineers in the development of a mammography retrieval system. The
application is built upon OutSystems, a low-code application platform. Key features
of the application include allowing experts to view probe images and associate them
with relevant images from the database. Additionally, the platform allows image filtering
based on eight mammogram dimensions. While the ultimate goal is to create
a system for medical specialists, the current platform represents a step in the process,
facilitating the acquisition of ground truth. As for the results obtained from the individual
models, in the training set, for the models of each dimension, they reach an
average accuracy of around 99.3%, while in the test set, the average accuracy is around
78%. Four approacheswere then developed for the final retrieval model, one assigning
equal weights to every dimension, another with empirically defined weights, a third
where the weights were defined according to the literature, and a final one where the
values of the weights were defined by a specialist. The quantitative results of the final
retrieval model according to the four approaches represent the similarity between
the probe image and the most similar image (the first image in the top-5). The similarity
results are the result of using the individual models in a weighted sum. The first
approach scored 0.319, the second 0.191, the third 0.197 and finally the last 0.292
Description
Keywords
Mammogram retrieval Breast cancer Image classification Deep learning Software