Related Experiment Video
Updated: Apr 21, 2026
13:44
Detection of Architectural Distortion in Prior Mammograms via Analysis of Oriented Patterns
Published on: August 30, 2013
43.9K
Size-Dependent Performance of Abnormal-Focused U-Net Segmentation for Mammographic Lesion Detection: A Two-Stage
Kian A Huang1, Haris K Choudhary1, Hailey Mangio1
1Radiology, University of South Florida Morsani College of Medicine, Tampa, USA.
Cureus
|April 20, 2026
Summary
A deep learning U-Net model effectively detected large breast cancer lesions but missed 27% of small abnormalities, highlighting a critical limitation for early cancer screening. Further development is needed for clinical deployment.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Breast cancer screening mammography faces challenges due to variable radiologist performance and missed cancers.
- Deep learning models show promise for automated lesion detection but often lack sufficient training data for abnormalities.
Purpose of the Study:
- To develop and evaluate a custom U-Net deep learning model for segmenting abnormal lesions in mammograms, focusing on improving detection of clinically relevant abnormalities.
- To address limitations in existing datasets biased towards normal cases by training on abnormal mammograms.
Main Methods:
- A custom U-Net architecture was trained using a two-stage approach on annotated abnormal mammograms from the Digital Mammography Dataset for Breast Cancer Diagnosis Research (DMID).
- Training involved patch-based pretraining and full-image fine-tuning with hybrid patch sampling, utilizing a composite loss function (focal and Tversky loss).
- Images were downsampled and contrast-enhanced (CLAHE) before evaluation using Dice coefficient, IoU, pixel accuracy, Hausdorff distance, and lesion detection rate, with size-stratified analysis.
Main Results:
- The model achieved a mean Dice score of 0.5793 and an overall lesion detection rate of 77.8% (43/55).
- Performance varied significantly by lesion size: large lesions had 100% detection (Dice 0.908), medium lesions 84.6% (Dice 0.724), and small lesions 73.0% (Dice 0.698).
- The primary failure was missed detection of small lesions (27% miss rate), likely due to information loss from downsampling high-resolution mammograms.
Conclusions:
- Abnormal-focused U-Net training demonstrates strong segmentation for large lesions but has critical limitations for small abnormalities, posing a significant barrier for early cancer detection.
- Downsampling high-resolution mammography to 224x224 pixels creates a resolution bottleneck, limiting the detection of subtle, small lesions.
- Clinical deployment is precluded by unknown false-positive rates on normal mammograms (absent from training) and requires future work on multi-scale architectures and hybrid training.
