Photometric Transformer Networks and Label Adjustment for Breast Density Prediction

Jaehwan Lee, Donggeun Yoo, Hyo-Eun KimWorkshops @ ICCV 2019 · DOI · arXiv

Breast density is a critical clinical factor in women’s health, serving as both a significant risk factor for breast cancer and a "masking" factor that complicates tumor detection on mammograms. Despite its importance, automating density grading is challenging due to inherent variability in medical imaging data. Researchers from Lunit Inc. have developed a deep learning framework utilizing Photometric Transformer Networks (PTN) and Label Distillation to address these challenges, significantly improving the accuracy and consistency of breast density classification.

The Challenges of Automated Density Assessment

Automating the assessment of breast density (typically categorized by BI-RADS grades $a$ through $d$) faces two primary obstacles:

  1. Intensity Variation (The Hardware Problem): Mammograms vary significantly based on machine manufacturer and hospital settings. Because density grading is intrinsically linked to pixel brightness, these inconsistencies can degrade the performance of standard AI models.
  2. Inter-reader Bias (The Human Problem): Radiologists often exhibit subjective differences when grading "borderline" cases. This "label noise" makes it difficult for models to learn clear, consistent decision boundaries between categories.

Methodology

To overcome these hurdles, the researchers introduced two key innovations designed to standardize both the input data and the training labels.

Photometric Transformer Networks (PTN)

Inspired by Spatial Transformer Networks, the PTN is a learnable module integrated at the front of the neural network. Unlike spatial transformers that handle geometric distortions, the PTN learns an image-specific piecewise linear function ($h$) to adjust the photometric properties—brightness and contrast—of the mammogram dynamically. This ensures that the classifier receives a normalized, consistent input regardless of the source hardware.

The PTN architecture learns a photometric transformation function h(x) to normalize input intensities before classification.

Label Distillation

To mitigate inconsistent human grading, a two-stage learning strategy was employed. A small "clean" subset of data ($D_s$), graded by a single expert specialist, was used to refine the labels in a much larger, noisier dataset ($D_r$). The process involves identifying samples where the model (trained on clean data) strongly disagrees with the original noisy label and then "blending" the original label with the model's prediction to create a more consistent "pseudo-label."

Density-AUC (dAUC)

Standard accuracy is often insufficient for density grading because the categories are ordinal ($a < b < c < d$). The authors proposed dAUC, which averages the Area Under the Curve across multiple binary splits (e.g., '$a$' vs. '$bcd$'). This provides a more nuanced evaluation of model performance across the density spectrum.

Experimental Results and Insights

The study utilized a massive dataset of 48,648 cases from five South Korean hospitals, with external validation on 8,206 cases from a US institution.

Learnable Normalization

The PTN proved superior to static normalization methods like CLAHE. By learning to normalize specifically for the task of density classification, the PTN reduced the standard deviation of image brightness across the dataset by over 60%, leading to much clearer visual inputs for the model.

Comparison between raw input mammograms and those normalized by the PTN, showing improved consistency in brightness and contrast.

Performance Gains

The proposed framework demonstrated dramatic improvements over baseline models:

  • Internal Accuracy: The full model achieved 79.4% accuracy, a significant leap from the 54.5% achieved by a standard ResNet-18 baseline.
  • External Generalization: On the US dataset, the model achieved a dAUC of 0.9481, outperforming the widely used clinical software LIBRA (0.8877).
  • Ablation Studies: Each component contributed to the final result, but the combination of PTN and Label Distillation provided the most substantial performance boost.

Conclusion

The integration of Photometric Transformer Networks and Label Distillation effectively addresses the core issues of hardware variability and reader subjectivity in medical imaging. By prioritizing data consistency over sheer quantity, this framework sets a new standard for robust breast density prediction. These techniques hold potential for broader application in other medical imaging tasks where hardware variation and label noise are prevalent.

Summary updated

BibTeX
@inproceedings{lee2019photometric,
  author = {Jaehwan Lee and Donggeun Yoo and Hyo-Eun Kim},
  title = {{Photometric Transformer Networks and Label Adjustment for Breast Density Prediction}},
  booktitle = {Workshops @ ICCV},
  year = {2019},
  doi = {10.1109/iccvw.2019.00058},
}
Figure