On the Use of Fully Convolutional Networks on Evaluation of Infrared Breast Image Segmentations

R. H. C. de Melo, Aura Conci, Cristina Nader Vasconcelos2017 · DOI

Abstract

Medical images usually must have their region of interest (ROI) segmented as a first step in a pattern recognition procedure. Automatic segmentation of these images is an open issue. This paper presents an automated technique to define the ROI for infrared breast exams, based on the use of Fully Convolutional Networks (FCN). Adequate comparison among new approaches by using available databases is very important, here some comparisons with other techniques are made. Moreover, concerning on line diagnosis, the comparison among possible techniques must be efficient enough to be done in real time. With our approach the time to segment the ROI was 100 milliseconds and the average accuracy obtained was 95%.

Introduction

In the field of medical imaging, identifying the Region of Interest (ROI) is a critical prerequisite for Computer-Aided Diagnosis (CAD) systems. For infrared breast thermography—a non-invasive and cost-effective diagnostic modality—manual ROI segmentation remains a significant bottleneck. The process is labor-intensive for specialists and prone to human subjectivity.

Automating this process requires a system that is both highly accurate and computationally efficient for real-time clinical applications. However, medical datasets are typically small, posing a challenge for deep learning models that often require large amounts of data to converge without overfitting.

Methodology

The researchers implemented a Fully Convolutional Network (FCN-8s) architecture, utilizing a VGG16 backbone. Unlike standard convolutional neural networks designed for classification, FCNs are capable of processing images of arbitrary dimensions and generating pixel-wise segmentation maps.

To address the constraints of a limited dataset, two primary strategies were employed:

  1. Transfer Learning: The network was initialized with weights pre-trained on the PASCAL VOC 2011 dataset. Despite being trained on natural images, the low-level features (e.g., edges and textures) proved highly effective for medical thermograms.
  2. Data Augmentation: The training set was expanded through horizontal mirroring—exploiting the natural symmetry of the human body—and spatial displacement, where 320x240 images were translated across a larger 500x500 canvas to ensure translation invariance.
Example of an original infrared breast thermogram used in the study.

Datasets and Training

The study utilized a public database from the Visual Lab at the Federal Fluminense University, consisting of 285 grayscale infrared images at a resolution of 320x240 pixels. Ground truth masks were manually segmented by specialists for comparison.

The corresponding ground truth segmentation mask for the infrared image.

A key finding was the necessity of pre-training; models initialized with random weights failed to converge, underscoring that transfer learning is essential for small-scale medical imaging tasks.

Results and Discussion

The proposed FCN approach demonstrated significant efficiency and robustness:

  • Accuracy: The model achieved an average pixel accuracy of 95.6%.
  • Performance: Inference time was recorded at 100 milliseconds, making it suitable for real-time "online" diagnosis.
  • Robustness: The FCN proved capable of handling "curious" samples, such as images where hair obscured the ROI. In some instances, the network correctly excluded hair from the ROI even when the human-provided ground truth did not, suggesting superior generalization.

Compared to traditional methods like Level Set Functions (LSF), which can reach 98% accuracy but require complex parameter tuning and pre-processing, the FCN offers an end-to-end solution with minimal intervention.

Conclusion

While the FCN approach is highly effective, it occasionally produces "little white islands"—small clusters of misidentified pixels. Future work will explore the integration of simple post-processing morphological filters to eliminate these artifacts, aiming to match the accuracy of traditional methods while maintaining superior processing speeds.

Summary updated

Abstract

Medical images usually must have their region of interest (ROI) segmented as a first step in a pattern recognition procedure. Automatic segmentation of these images is an open issue. This paper presents an automated technique to define the ROI for infrared breast exams, based on the use of Fully Convolutional Networks (FCN). Adequate comparison among new approaches by using available databases is very important, here some comparisons with other techniques are made. Moreover, concerning on line diagnosis, the comparison among possible techniques must be efficient enough to be done in real time. With our approach the time to segment the ROI was 100 milliseconds and the average accuracy obtained was 95%.
BibTeX
@misc{melo2017use,
  author = {R. H. C. de Melo and Aura Conci and Cristina Nader Vasconcelos},
  title = {{On the Use of Fully Convolutional Networks on Evaluation of Infrared Breast Image Segmentations}},
  year = {2017},
  doi = {10.5753/sbcas.2017.3701},
}
Figure