Breast cancer annually affects approximately 2.3 million women worldwide
and results in more than 1,300 new diagnoses in Slovenia. Despite their
effectiveness, screening tests are limited by the large volume of data, interobserver
variability, and false results. Although artificial intelligence methods
for prediction exist, they often fail to establish sufficient trust among radiologists.
This master’s thesis presents the design of explanations of breast cancer
predictions for radiologists based on the segmentation of mammographic images.
After an initial comparison of three models (UNeXt, SegFormer, and
MedSAM), we selected the pre-trained MedSAM model, which we fine-tuned
using the CSAW-CC dataset. Predictions were performed on image patches,
followed by reconstruction of the full mammographic images, with lesions
on the final image clearly represented by outlining. On the test set, a Dice
coefficient of 0.72 and an IoU of 0.59 were achieved, which is comparable to
the literature. In the survey, radiologists confirmed their satisfaction with
the predictions. The work contributes to the design and development of
explainable tools that are primarily intended for end users – radiologists.
|
