U-Net Based Medical Image Segmentation on Breast Cancer Mammograms
Date Issued
2025-04
Author(s)
Hajisoy, Minura
Abstract
Breast cancer is still one of the main causes of cancer related mortality for the women throughout the
world, where almost 670 000 reported death and over 2.3 million new cases in 2022. For improving the outcomes
and reducing the treatment costs for patients, early detection of unhealthy lesions in breast medical images is quite
important. In this study, using the Curated Breast Imaging Subset of the Digital Database for Screening
Mammography (CBIS-DDSM) dataset a comprehensive deep-learning based segmentation pipeline is created to
evaluate and solve challenges faced when working on high-resolution medical images, in this case breast
mammograms with the help of U-Net deep learning architecture.
Starting with CBIS-DDSM dataset, the first addressed issue has been general preprocessing issues in the
dataset such as combining full mammograms and region-of-interest (ROI) masks with the help of automated
dictionary mapping for preparing them for augmentation. Augmentation step is done by cropping image mask
pairs into n × n grids (for n ranging from 2 to 16) with the shape 256 × 256. These pairs are then fed into a U-Net
model optimized with a composite loss function which is a combination of binary cross-entropy and Dice loss.
During the inference, those patch-level probability maps are combined into a full size mammogram
images and masks, and comprehensive pixel-count thresholding techniques such as intersection and union of them
are employed. For each of the images inside the test set, the threshold value t which maximizes the Intersection
over Union (IoU) is selected and compared to the original ROI mask of that exact image. Final achieved metrics
values are: average IoU value of 0.60, average accuracy of 0.99, precision value of 0.72, recall value of 0.77, and
finally F1 score of 0.71. According to those qualitative results, the model reliably detects the lesion boundaries
even in some challenging scenarios where the mammogram have some dense tissues. It is noteworthy to highlight
that cropping-based U-Net architecture might have a potential to work and imrpove with different mechanisms in
the future.
world, where almost 670 000 reported death and over 2.3 million new cases in 2022. For improving the outcomes
and reducing the treatment costs for patients, early detection of unhealthy lesions in breast medical images is quite
important. In this study, using the Curated Breast Imaging Subset of the Digital Database for Screening
Mammography (CBIS-DDSM) dataset a comprehensive deep-learning based segmentation pipeline is created to
evaluate and solve challenges faced when working on high-resolution medical images, in this case breast
mammograms with the help of U-Net deep learning architecture.
Starting with CBIS-DDSM dataset, the first addressed issue has been general preprocessing issues in the
dataset such as combining full mammograms and region-of-interest (ROI) masks with the help of automated
dictionary mapping for preparing them for augmentation. Augmentation step is done by cropping image mask
pairs into n × n grids (for n ranging from 2 to 16) with the shape 256 × 256. These pairs are then fed into a U-Net
model optimized with a composite loss function which is a combination of binary cross-entropy and Dice loss.
During the inference, those patch-level probability maps are combined into a full size mammogram
images and masks, and comprehensive pixel-count thresholding techniques such as intersection and union of them
are employed. For each of the images inside the test set, the threshold value t which maximizes the Intersection
over Union (IoU) is selected and compared to the original ROI mask of that exact image. Final achieved metrics
values are: average IoU value of 0.60, average accuracy of 0.99, precision value of 0.72, recall value of 0.77, and
finally F1 score of 0.71. According to those qualitative results, the model reliably detects the lesion boundaries
even in some challenging scenarios where the mammogram have some dense tissues. It is noteworthy to highlight
that cropping-based U-Net architecture might have a potential to work and imrpove with different mechanisms in
the future.