ROI Extraction in Thermographic Breast Images Using Genetic Algorithms

L. C. Mendes, Érick Oliveira Rodrigues, Sandro C. Izidoro, Aura Conci, Panos LiatsisIWSSIP 2020 · OpenReview

TL;DR

This paper introduces a novel, fully automatic method for extracting the Region of Interest (ROI) in thermographic breast images using Genetic Algorithms (GAs) and a cardioid-shaped fitness function. This approach aims to improve the accuracy of breast cancer detection by isolating the relevant breast area without requiring manual input.

What Problem Does This Paper Address?

Accurate identification of the breast region in thermographic images is crucial for reliable breast cancer detection. Existing methods often require manual intervention, such as seed point selection, which can be time-consuming and introduce variability. Furthermore, non-breast related information in the image can negatively impact the accuracy of diagnostic systems. This paper addresses the need for an automated, robust, and accurate ROI extraction method that can handle variations in thermographic images, including those from both healthy and diseased breasts.

Methods

The core of this work lies in the application of Genetic Algorithms (GAs) for ROI extraction. The authors propose a novel fitness function that utilizes the shape of a cardioid. A cardioid is a heart-shaped curve defined by a polar equation, and its base parameter is evolved by the GA to fit the breast region. The fitness function evaluates how well the cardioid encloses the breast area by considering the grayscale intensity of pixels within and outside the cardioid. Specifically, the image is divided into four grayscale groups: Group A (background), Group D (irrelevant body areas like arms/belly), and Groups B and C (containing chest, abdomen, and the breast area of interest, respectively). The fitness function assigns weights to these groups, penalizing the inclusion of Group A and D, and rewarding the inclusion of Group C.

The GA population consists of individuals, each representing a potential cardioid defined by its x-position, y-position, and base size. The algorithm iteratively evolves this population through selection (tournament selection), crossover (one-point crossover), and mutation. The use of gray coding for the cardioid parameters is highlighted as a way to mitigate the "Hamming cliff problem." The process is fully automatic, requiring no manual seed points.

Datasets

The study utilized a dataset of 58 infrared breast images from the Database for Mammography Research with Infrared Images (DMR-IR). These images were captured using a FLIR ThermaCam S45 camera and included both healthy and sick breasts (those with breast diseases like cancer). The dataset was collected from patients older than 35 years and corrected for environmental factors.

Key Insights and Ideas

  • Cardioid as a Shape Prior: Employing a cardioid shape as a prior for the breast region is a novel approach. This shape is intuitively related to the breast's form and allows for a parameterized representation that GAs can effectively optimize.
  • Fully Automatic ROI Extraction: The elimination of manual seed point selection is a significant advancement, making the method more practical and scalable for clinical applications.
  • Fitness Function Design: The clever design of the fitness function, incorporating grayscale information and weighted penalties for irrelevant areas, is key to the algorithm's success in distinguishing the breast from the background and other body parts.
  • GA's Suitability for Image Segmentation: This work demonstrates the power of GAs in tackling complex image segmentation problems, particularly when dealing with noisy or variable data like thermographic images.
  • Impact of Breast Health on Homogeneity: The study notes that healthy breasts exhibit more homogeneous gray-level intensity patterns, leading to better ROI extraction results compared to sick breasts, which can have more heterogeneous patterns due to disease progression.

Key Results

The proposed GA-based method achieved successful separation of the breast region in 52 out of 58 images. Perfect ROI extraction was achieved in 16 cases, with satisfactory results in an additional 36 cases. The remaining 6 cases showed poorer results, where the cardioid was displaced. The study also observed that the GA showed significant improvement in fitness scores within the first 50 generations, with a slower increase thereafter. The processing time increased linearly with the number of generations.

Limitations and Future Work

The authors acknowledge that the initial approach using two independent ellipses for each breast was problematic, especially for sick patients, due to temperature discrepancies and the interference of a black metallic marker used during image acquisition. While the cardioid formulation improved this, the marker still posed a challenge, causing the cardioid to sometimes drift away from the breast area.

Future work suggestions include:

  • Incorporating Texture Information: Enhancing the fitness function with texture features (e.g., from co-occurrence matrices, Laplacian filters) could improve robustness.
  • Improving GA Parameters and Fitness Function: Further optimization of GA parameters and the fitness function is suggested.
  • Addressing the Black Marker: Investigating methods to mitigate the impact of the black marker, such as replacing it with random gray information, is proposed.
  • Exploring Affine Transforms: Applying affine transforms (rotation, warps, distortions) to the cardioid shape to better adapt to different breast types.
  • Investigating Hybrid Approaches: Exploring combinations of GA with fuzzy logic or neural networks (GA+Fuzzy or GA+RNA) for further improvements.

Summary updated

Abstract

This work proposes the use of Genetic Algorithms (GA) to identify the area of the breast from the background in thermographic breast images. The proposed method uses color information, a fitness function based on cardioids, and GA. This is the first work in the literature to propose a Region of Interest (ROI) extraction based on GA and cariods. ROI extraction can improve the accuracy of cancer detection and assist with the standardization of acquisition protocols. The method is able to successfully separate the breast region in 52 out of 58 images, while being fully automatic, and not requiring manual selection of seed points.
BibTeX
@inproceedings{mendes2020roi,
  author = {L. C. Mendes and Érick Oliveira Rodrigues and Sandro C. Izidoro and Aura Conci and Panos Liatsis},
  title = {{ROI Extraction in Thermographic Breast Images Using Genetic Algorithms}},
  booktitle = {IWSSIP},
  year = {2020},
  url = {https://openreview.net/forum?id=yXzFHcRnjw},
}
Figure