A New Automatic Change Detection Frame-work Based on Region Growing and Weighted Local Mutual Information: Analysis of Breast Tumor Response to Chemotherapy in Serial MR Images

Narges Norouzi, Reza Azmi, Nooshin Noshiri, Robab AnbiaeearXiv 2021 · arXiv

Abstract

Monitoring how a breast tumor responds to chemotherapy requires pinpointing subtle changes across serial MRIs—a task that is notoriously difficult for human eyes and traditional algorithms alike. This paper introduces a new automated framework that combines Weighted Local Mutual Information (WLMI) with a region-growing strategy to detect these evolutions with higher precision and better noise resistance than previous state-of-the-art methods.

What Problem Does This Paper Address?

When a patient undergoes neoadjuvant chemotherapy for breast cancer, clinicians must accurately assess whether the tumor is shrinking to adjust treatment plans. Magnetic Resonance Imaging (MRI) is the gold standard for this monitoring, but comparing "before" and "after" scans presents significant challenges.

The difficulties are primarily three-fold:

  1. Tissue Complexity: Breast tissue is soft and deforms easily between scans, making perfect alignment (registration) nearly impossible.
  2. Imaging Artifacts: MRI scans often suffer from Rician noise and intensity inhomogeneities.
  3. Subtle Evolutions: Early responses to treatment are often too faint for manual detection by radiologists.
Taxonomy of image comparison approaches, highlighting the path of intensity-based probabilistic methods.

Existing automated tools, such as the Generalized Likelihood Ratio Test (GLRT), often fail because they rely on simple intensity averages, which are easily compromised by noise or slight misalignments. As shown in the taxonomy above, the proposed method seeks to improve upon intensity-based probabilistic approaches.

Methods

The authors propose a two-phase pipeline designed to filter out imaging noise and focus purely on biological change.

The proposed two-phase framework for automatic change detection in breast MR images.

Phase 1: Preprocessing & EROC

Beyond standard noise reduction and B-spline registration, the authors introduce Hierarchical Histogram Matching (HHM) for intensity normalization, which is more robust than linear methods. A critical component is the Extraction Region of Changes (EROC). Instead of analyzing the entire image, EROC uses similarity curves and a modified hill-climbing algorithm to identify specific "zones of interest" where significant changes are likely occurring. This step drastically reduces computation time and minimizes false positives.

Phase 2: Detection via WLMI and Region Growing

The core innovation is Weighted Local Mutual Information (WLMI). While standard Mutual Information measures shared information between two images, the "Weighted" version assigns varying importance to pixels in a neighborhood based on their distance from the center and the intensity of the change. This allows the algorithm to ignore random noise while amplifying subtle, consistent signals of tumor evolution.

Finally, a Region Growing algorithm ensures "local consistency." Since biological changes typically occur in clusters rather than isolated pixels, this step prunes away "forged" changes (like single-pixel noise) that do not belong to a larger evolving structure.

Experimental Setup

The framework was validated using two primary datasets:

  • Simulated Data: Using the PIDER breast MRI dataset, the authors simulated 27 images with tumor volume changes ranging from 5% to 90% to establish a "ground truth" for accuracy measurement.
  • Real Clinical Data: 30 longitudinal MR images from patients at Imam Reza Hospital, representing real-world chemotherapy cycles. Evaluation relied on expert radiologist labeling to determine True Positives and False Positives.

Key Results and Insights

The proposed framework consistently outperformed the standard GLRT method across multiple metrics:

  • Precision Boost: On real clinical images, the WLMI framework achieved an average precision of 76.13%, a significant 16% improvement over GLRT (60.27%).
  • Subtle Change Detection: For small changes (<10% tumor volume), the framework maintained a much higher Area Under the Curve (Az) of 93.49 compared to GLRT’s 63.87.
  • Efficiency: The system can scan 27 images in roughly 81 seconds, whereas a human expert typically requires about 15 minutes for a similar comparison.

The study highlights that weighting central pixels more heavily allows WLMI to overcome the "blurring" effect of average-intensity methods. Furthermore, the algorithm was able to detect subtle evolutions that human experts frequently missed.

Limitations and Future Work

Despite its strengths, the system has areas for improvement:

  • De Novo Detection: The EROC method is currently tuned for evolution rather than detecting brand-new tumors.
  • Microcalcifications: The system is currently inefficient at detecting microcalcifications, which are often indistinguishable from image noise.
  • Automation of Thresholds: The EROC phase still requires experimental estimation of threshold values. Future work may involve more advanced "Demons" algorithms or supervised model-based methods to further automate this process.

Summary updated

BibTeX
@misc{norouzi2021new,
  author = {Narges Norouzi and Reza Azmi and Nooshin Noshiri and Robab Anbiaee},
  title = {{A New Automatic Change Detection Frame-work Based on Region Growing and Weighted Local Mutual Information: Analysis of Breast Tumor Response to Chemotherapy in Serial MR Images}},
  year = {2021},
  eprint = {2110.10242},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2110.10242},
}
Figure