Breast Mass Detection in Mammography Based on Image Template Matching and CNN

Lilei Sun, Huijie Sun, Junqian Wang, Shuai Wu, Yong Zhao, Yong XuSensors 2021 · DOI

TL;DR

This paper introduces a novel, multi-stage approach for detecting breast masses in mammograms by mimicking human visual inspection. By combining mathematical morphology for initial feature enhancement, image template matching for candidate region identification, a Convolutional Neural Network (CNN) for classification, and Particle Swarm Optimization (PSO) for bounding box refinement, the method achieves superior detection performance compared to existing techniques.

What Problem Does This Paper Address?

Breast cancer remains a significant global health concern, and early detection through mammography is crucial for improving patient outcomes. While automated breast mass detection systems aim to assist radiologists, current methods often struggle with limited accuracy, particularly due to the inherent challenges in mammographic images: blurry edges, complex textures, and overlapping tissues. This can lead to missed diagnoses or false alarms. The paper addresses the need for a more robust and accurate breast mass detection system that can effectively overcome these challenges and better emulate the nuanced visual search and identification process employed by human experts.

Methods

The proposed method employs a four-phase pipeline:

  1. Mammographic Image Processing (Mathematical Morphology): This initial step aims to enhance the visibility of potential breast masses. It involves an "eroding" operation to remove noise and highlight high-intensity regions, followed by a "dilating" operation to normalize these regions into approximate circular shapes. This pre-processing is designed to make subsequent matching more effective.
  2. Candidate Region Generation (Image Template Matching): A custom circular image template, derived from a typical breast mass, is used to scan the processed mammograms. Regions exhibiting high similarity to this template are identified as "suspected regions of breast mass." This step leverages the idea that breast masses often have a somewhat circular appearance.
  3. Candidate Region Identification (BD-CNN): A Convolutional Neural Network (CNN), termed BD-CNN, is employed to classify the suspected regions identified in the previous step. This network, composed of convolutional blocks, max-pooling, batch normalization, and fully connected layers, distinguishes between actual breast masses and background tissue.
  4. Bounding Box Regression (PSO): The bounding boxes generated by the initial template matching are often rough. To refine these, the Particle Swarm Optimization (PSO) algorithm is used. PSO searches for an optimal location and bounding box that best matches the ground truth, using the BD-CNN's classification probability as its fitness function. This iterative optimization process aims to achieve more precise localization of the detected masses.

Datasets

The study utilizes the Digital Database for Screening Mammography (DDSM) dataset for both training and evaluation. A subset of 439 mammographic images was selected. The dataset was split into training (70%) and testing (30%) sets. The authors mention that the training and test datasets contain 307 and 132 mammographic images, respectively.

Key Insights and Ideas

  • Mimicking Human Vision: The core innovation lies in simulating the human visual search strategy. By first enhancing potential regions and then using a template matching approach, the method aims to replicate how radiologists might initially scan an image for suspicious areas.
  • Multi-Stage Synergy: The paper demonstrates the power of combining different techniques. Mathematical morphology prepares the image, template matching provides initial candidates, CNN offers robust classification, and PSO refines the localization, with each stage building upon the strengths of the others.
  • PSO for Bounding Box Refinement: Using PSO to optimize bounding box regression, guided by a CNN's confidence score, is a clever way to improve localization accuracy beyond what initial detection methods provide. This addresses the common issue of imprecise bounding boxes.
  • Circular Template Effectiveness: The use of a circular template, while seemingly simple, is justified by the typical shape of breast masses and its effectiveness in conjunction with the morphological pre-processing.

Key Results

The proposed method was evaluated on the DDSM dataset and demonstrated superior performance compared to several state-of-the-art breast mass detection methods. Specifically:

  • The method achieved the highest True Positive Rate (TPR) and the lowest False Positives Per Image (FPI) among the compared methods.
  • In Table 1, the proposed method achieved a Deep learning TPR of 96%, Accuracy of 85.82%, Precision of 50.81%, Recall of 95.38%, F1 Score of 66.31%, and an FPI of 0.53. This is notably better than many traditional and deep learning approaches listed.
  • The ROC curve analysis (Figure 10) shows an AUC of 0.89 for the proposed method, outperforming RetinaNet (0.81) and FSAF (0.83).
  • Figure 12 illustrates that higher classification probability from BD-CNN correlates with a higher Intersection over Union (IoU) for bounding box regression, indicating the effectiveness of the PSO refinement.

Limitations and Future Work

  • Shape Limitations: The paper acknowledges that the image template matching method may struggle to match breast masses that are not circular or elliptical in shape, as illustrated in Figure 11.
  • Threshold Sensitivity: The choice of threshold for image template matching significantly influences the number of suspected regions and, consequently, the detection results (Figure 9).
  • Future Work: While not explicitly detailed, the success of this multi-stage approach suggests that further exploration of different morphological operations, template designs, and optimization algorithms could lead to even greater improvements. Expanding the dataset and testing on diverse populations could also be valuable.

Summary updated

BibTeX
@article{sun2021breast,
  author = {Lilei Sun and Huijie Sun and Junqian Wang and Shuai Wu and Yong Zhao and Yong Xu},
  title = {{Breast Mass Detection in Mammography Based on Image Template Matching and CNN}},
  journal = {Sensors},
  year = {2021},
  doi = {10.3390/s21082855},
}
Figure