Slice-Aware Attention for Breast MRI Segmentation Quality Control

In modern medical imaging, deep learning models are becoming increasingly capable of segmenting tumors in MRI scans automatically. But as these models are deployed at scale, a new challenge emerges: how do we know when their segmentations are reliable?

Manual quality control (QC) by radiologists is time-consuming, subjective, and simply doesn’t scale to thousands of scans. To address this, I set out to develop an automated quality control framework that can flag when a segmentation likely needs correction. This work was conducted during the BigDataX REU at the University of Chicago, in collaboration with the Data Science Institute and supported by the NSF.

The Dataset: MAMA-MIA

I used the MAMA-MIA dataset — a multi-center collection of over 1,500 pre-treatment breast DCE-MRI scans, each with expert-corrected tumor segmentations. Radiologists rated each segmentation as Good, Acceptable, Poor, or Missed.

For simplicity, I grouped them into two classes:

• Good: both experts marked it as good
• Bad: all other cases

The dataset covers diverse imaging protocols across several institutions, making it an excellent benchmark for real-world robustness.

Preprocessing Pipeline

Medical data preprocessing turned out to be as important as model design. Each MRI scan went through the following steps:

1. N4 Bias Correction – removes low-frequency intensity artifacts
2. Denoising – reduces noise while preserving edges
3. Resampling – ensures uniform voxel spacing across scans
4. Cropping and Padding – isolates single-breast volumes, resized to 256×256×D
5. Intensity Normalization – scales pixel intensities for stable training
6. Filtering – excluded sagittal and double-breast scans