Automatic Detection and Segmentation of Thrombi in Abdominal Aortic Aneurysms Using a Mask Region-Based Convolutional Neural Network with Optimized Loss Functions

Byunghoon Hwang, Jihu Kim, Sungmin Lee, Eunyoung Kim, Jeongho Kim, Younhyun Jung, Hyoseok Hwang

TL;DR

This paper introduces a novel deep learning approach for automatically detecting and segmenting thrombi within abdominal aortic aneurysms (AAAs) using an enhanced Mask R-CNN. By optimizing loss functions for both bounding box regression and segmentation, the method achieves superior accuracy and efficiency, crucial for timely patient management and reducing mortality associated with AAAs.

What Problem Does This Paper Address?

Abdominal aortic aneurysms (AAAs) are a significant cause of mortality, often due to rupture. Monitoring AAA progression and thrombus formation is critical for patient care, with treatments like open or endovascular repair chosen based on detailed analysis. Computed tomography angiography (CTA) is the preferred imaging method, but manual detection and segmentation of thrombi from CTA images are challenging. Thrombi have irregular shapes, similar intensity to surrounding tissues, and can be obscured by medical devices like stent grafts. This manual process is time-consuming, prone to errors, and data is limited, highlighting the need for automated, accurate solutions.

Methods

The core of this work is an adaptation of the Mask R-CNN framework, a popular deep learning model for object detection and segmentation. The authors propose two key innovations:

  • Optimized Bounding Box Regression Loss (SCIoU): They introduce a new loss function called SCIoU, which combines the strengths of the smooth L1 loss (for fast convergence) and the Complete Intersection over Union (CIoU) loss (for accurate bounding box prediction considering overlap, center distance, and aspect ratio). This hybrid approach aims to improve the precision of detecting the thrombus's location.
  • Weighted Binary Focal Loss for Segmentation: To enhance the accuracy of segmenting the thrombus, the standard cross-entropy loss is replaced with a weighted binary focal loss. This modification focuses more on difficult-to-classify examples and reduces mis-segmentation, particularly for irregular thrombus shapes.

The model was trained using 2D slices from 3D CTA datasets.

Datasets

The study utilized a dataset of AAA thrombus CTA scan images from 60 unique patients, collected between 2012 and 2020 at Gachon University Gil Medical Center. This dataset is noted as being one of the largest post-operation AAA thrombus datasets available, contributing to the robustness of the proposed method. The CTA images were acquired using five different Siemens scanning equipment models. Ground truth labels for thrombi were manually created by trained radiologists.

Key Insights and Ideas

  • The Importance of Joint Detection and Segmentation: The paper emphasizes that for clinical relevance, both detecting the presence of a thrombus and accurately segmenting its boundaries are crucial, rather than focusing on just one aspect.
  • Loss Function Engineering is Key: The authors demonstrate that carefully designing and combining loss functions can significantly improve the performance of existing deep learning architectures like Mask R-CNN for specific medical imaging tasks.
  • SCIoU Balances Speed and Accuracy: The novel SCIoU loss function effectively balances the need for fast convergence with precise bounding box regression, outperforming individual loss functions.
  • Focal Loss for Better Segmentation: Adapting focal loss for binary segmentation proves effective in handling the challenging visual characteristics of thrombi in CTA scans.

Key Results

The proposed method achieved impressive results:

  • Thrombus Detection: The SCIoU loss function yielded the highest F1 score for thrombus detection at 0.9197, with a precision of 0.8847 and recall of 0.9576. This outperformed other tested bounding box regression loss functions and existing methods like DeepAAA and DetectNet.
  • Thrombus Segmentation: The combined approach (CL Mask R-CNN) achieved the highest segmentation performance across several metrics, including an overlap score of 0.8971, Jaccard index of 0.7163, and Dice coefficient of 0.8267. Notably, it achieved the lowest false negative rate of 0.1029, which is critical for clinical decision-making.

Qualitative results showed significant reductions in false negatives compared to other methods.

Limitations and Future Work

  • Limitations: The paper acknowledges that while the proposed method significantly reduces false negatives, false positives can increase in certain areas, though this is considered less critical than false negatives in a clinical context. The dataset, while large, is from a single institution.
  • Future Work: The authors plan to explore more advanced instance segmentation models, such as Mask Scoring R-CNN, to further enhance accuracy for thrombus detection and segmentation. They also aim to leverage these improved models as backbone networks for future research.

Summary updated

Abstract

The detection and segmentation of thrombi are essential for monitoring the disease progression of abdominal aortic aneurysms (AAAs) and for patient care and management. As they have inherent capabilities to learn complex features, deep convolutional neural networks (CNNs) have been recently introduced to improve thrombus detection and segmentation. However, investigations into the use of CNN methods is in the early stages and most of the existing methods are heavily concerned with the segmentation of thrombi, which only works after they have been detected. In this work, we propose a fully automated method for the whole process of the detection and segmentation of thrombi, which is based on a well-established mask region-based convolutional neural network (Mask R-CNN) framework that we improve with optimized loss functions. The combined use of complete intersection over union (CIoU) and smooth L1 loss was designed for accurate thrombus detection and then thrombus segmentation was improved with a modified focal loss. We evaluated our method against 60 clinically approved patient studies (i.e., computed tomography angiography (CTA) image volume data) by conducting 4-fold cross-validation. The results of comparisons to multiple other state-of-the-art methods suggested the superior performance of our method, which achieved the highest F1 score for thrombus detection (0.9197) and outperformed most metrics for thrombus segmentation.
BibTeX
@article{hwang2022automatic,
  author = {Byunghoon Hwang and Jihu Kim and Sungmin Lee and Eunyoung Kim and Jeongho Kim and Younhyun Jung and Hyoseok Hwang},
  title = {{Automatic Detection and Segmentation of Thrombi in Abdominal Aortic Aneurysms Using a Mask Region-Based Convolutional Neural Network with Optimized Loss Functions}},
  journal = {Sensors},
  year = {2022},
  doi = {10.3390/s22103643},
}
Figure