Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning using privileged information for segmenting tumors on digital mammograms

Ioannis N. Tzortzis, Konstantinos Makantasis, Ioannis Rallis, Nikolaos Bakalos, Anastasios Doulamis, Nikolaos Doulamis

TL;DR

This work tackles data scarcity and GDPR-driven limits in mammogram tumor segmentation by applying Learning Using Privileged Information (LUPI). A teacher network learns from privileged, enhanced patches while a student network trains on original patches, with the student guided by teacher predictions via a combined loss $L_{PI}$. Experiments on the INBreast-derived patch dataset show the privileged-information approach often yields up to 10% higher F1 scores and lower variance compared to a baseline, demonstrating data-efficient and robust segmentation. The findings suggest practical impact for privacy-preserving training and potential integration with federated learning to improve distributed medical image analysis.

Abstract

Limited amount of data and data sharing restrictions, due to GDPR compliance, constitute two common factors leading to reduced availability and accessibility when referring to medical data. To tackle these issues, we introduce the technique of Learning Using Privileged Information. Aiming to substantiate the idea, we attempt to build a robust model that improves the segmentation quality of tumors on digital mammograms, by gaining privileged information knowledge during the training procedure. Towards this direction, a baseline model, called student, is trained on patches extracted from the original mammograms, while an auxiliary model with the same architecture, called teacher, is trained on the corresponding enhanced patches accessing, in this way, privileged information. We repeat the student training procedure by providing the assistance of the teacher model this time. According to the experimental results, it seems that the proposed methodology performs better in the most of the cases and it can achieve 10% higher F1 score in comparison with the baseline.

Learning using privileged information for segmenting tumors on digital mammograms

TL;DR

This work tackles data scarcity and GDPR-driven limits in mammogram tumor segmentation by applying Learning Using Privileged Information (LUPI). A teacher network learns from privileged, enhanced patches while a student network trains on original patches, with the student guided by teacher predictions via a combined loss . Experiments on the INBreast-derived patch dataset show the privileged-information approach often yields up to 10% higher F1 scores and lower variance compared to a baseline, demonstrating data-efficient and robust segmentation. The findings suggest practical impact for privacy-preserving training and potential integration with federated learning to improve distributed medical image analysis.

Abstract

Limited amount of data and data sharing restrictions, due to GDPR compliance, constitute two common factors leading to reduced availability and accessibility when referring to medical data. To tackle these issues, we introduce the technique of Learning Using Privileged Information. Aiming to substantiate the idea, we attempt to build a robust model that improves the segmentation quality of tumors on digital mammograms, by gaining privileged information knowledge during the training procedure. Towards this direction, a baseline model, called student, is trained on patches extracted from the original mammograms, while an auxiliary model with the same architecture, called teacher, is trained on the corresponding enhanced patches accessing, in this way, privileged information. We repeat the student training procedure by providing the assistance of the teacher model this time. According to the experimental results, it seems that the proposed methodology performs better in the most of the cases and it can achieve 10% higher F1 score in comparison with the baseline.
Paper Structure (11 sections, 4 equations, 3 figures, 2 tables)

This paper contains 11 sections, 4 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Methodology
  4. Problem formulation
  5. The segmentation model architecture
  6. Experimental results
  7. Dataset description
  8. Data pre-processing
  9. Experimental setup
  10. Performance evaluation
  11. Conclusion

Figures (3)

  • Figure 1: a) The original mammogram, b) the output after applying histogram equalization, c) the output after applying contrast adjustment, d) the corresponding annotation mask.
  • Figure 2: The UNet model
  • Figure 3: The experimental results grouped by the amount of samples utilized for the training of the models. Each plot shows the performance of all the models which is measured using the F1 score.