Automated Whole-Body Bone Lesion Detection for Multiple Myeloma on 68Ga-Pentixafor PET/CT Imaging Using Deep Learning Methods

Lina Xu, Giles Tetteh, Jana Lipková, Yu Zhao, Hongwei Li, Patrick Ferdinand Christ, Marie Piraud, Andreas K. Buck, Kuangyu Shi, Bjoern MenzeContrast 2018 · DOI

TL;DR

This paper introduces a novel deep learning approach, W-Net, for automated whole-body bone lesion detection in Multiple Myeloma (MM) patients using <sup>68</sup>Ga-Pentixafor PET/CT scans. By effectively combining anatomical (CT) and molecular (PET) information, W-Net significantly outperforms traditional machine learning methods, paving the way for more efficient and accurate MM diagnosis.

What Problem Does This Paper Address?

Multiple Myeloma (MM) is a blood cancer characterized by the proliferation of malignant plasma cells, often leading to bone lesions. Accurately identifying and characterizing these bone lesions is crucial for diagnosis, staging, and treatment monitoring. While advanced imaging techniques like PET/CT offer improved visualization, the sheer number and heterogeneity of lesions across the entire body make manual detection a tedious, time-consuming, and error-prone process, leading to significant inter-observer variability. Existing computer-aided detection (CAD) methods have limitations in handling the complex, multimodal nature of PET/CT data and the inherent variability of MM lesions. This paper aims to bridge this gap by developing an automated, deep learning-based system for robust whole-body bone lesion detection.

Methods

The core of this work lies in the application and adaptation of deep learning architectures for 3D volumetric image segmentation and detection. The researchers employed two convolutional neural networks (CNNs):

  • V-Net: A well-established 3D CNN architecture designed for volumetric segmentation. It utilizes a contracting path for feature extraction and an expanding path for reconstruction, allowing it to learn anatomical knowledge from CT data and then incorporate PET/CT information for lesion prediction.
  • W-Net: This is a novel architecture proposed by the authors, built upon V-Net. It cascades two V-Nets in a "W" shape. The first V-Net focuses on learning anatomical information from CT to generate a binary skeleton mask of the bone. This mask is then fed into the second V-Net along with the PET/CT data. This hierarchical approach allows the network to first understand the skeletal structure and then use this as a form of regularization to improve lesion detection, especially for smaller or less distinct lesions.

The W-Net architecture aims to leverage the complementary information from PET and CT to achieve more accurate segmentation and detection. The training process involved class balancing techniques to address the imbalance between lesion and non-lesion data and a patch-based strategy to improve the representation of small lesions.

Datasets

The study utilized two types of datasets:

  • Digital Phantoms: Realistic digital phantoms of <sup>68</sup>Ga-Pentixafor PET scans were generated using physically based simulations. These phantoms provided a ground truth for evaluating the performance of the deep learning algorithms. A total of 70 digital phantoms with simulated lesions of varying sizes, shapes, and uptakes were created.
  • Clinical Data: 12 patients with histologically proven MM who underwent <sup>68</sup>Ga-Pentixafor PET/CT imaging were included. For these patients, lesions were manually annotated by experienced radiologists under supervision, serving as the ground truth for evaluating the models on real-world data.

Key Insights and Ideas

  • Multimodal Fusion is Key: The study strongly emphasizes that combining PET and CT data is crucial for improving MM bone lesion detection. While CT provides anatomical context, PET offers molecular information, and their synergistic use leads to better lesion characterization.
  • Hierarchical Learning for Structure and Detail: The W-Net architecture's success highlights the benefit of a hierarchical approach. By first learning the bone structure (skeleton mask) and then using this as a constraint for lesion detection, the network can better handle the complex background and improve accuracy.
  • Deep Learning Outperforms Traditional Methods: The results clearly demonstrate that deep learning approaches, particularly W-Net, significantly outperform traditional machine learning methods (Random Forest, k-NN, SVM) in terms of Dice score, indicating superior segmentation accuracy.
  • Handling Data Imbalance: The use of class balancing loss functions and patch-based sampling strategies was essential for training deep learning models on datasets with a high imbalance between lesion and non-lesion data, a common challenge in medical imaging.

Key Results

  • Phantom Study: The V-Net model achieved a high specificity of 99.68% and a Dice score of 89.26% on the simulated digital phantoms, demonstrating the potential of deep learning for whole-body MM lesion detection.
  • Clinical Data Comparison:
    • Combining PET and CT in V-Net significantly improved the Dice score (69.49%) compared to using CT alone (26.37%) or PET alone (28.51%).
  • The proposed W-Net architecture, using PET/CT input, achieved the highest segmentation accuracy with a Dice score of 72.98%. It also demonstrated superior sensitivity (73.50%) and specificity (99.59%) compared to V-Net with PET/CT.
  • Traditional Methods: While traditional methods showed comparable sensitivity and specificity, their Dice scores were significantly lower (ranging from 21.69% to 26.94%), underscoring the advantage of deep learning for accurate segmentation.

Limitations and Future Work

  • Limited Patient Data: The study acknowledges that the current dataset of 12 clinical patients is small, which may have hampered the performance of the deep learning models. Augmenting the training data through patch generation helped, but more data is needed for further improvement.
  • Focus on Bone Lesions: The current W-Net architecture is specifically designed for bone lesions. Future work should extend this to include extramedullary lesions, which are common in MM patients, by incorporating more labels for different tissue types.
  • Computational Cost: The W-Net architecture, with its cascaded V-Nets, incurs higher computational costs. Further optimization may be necessary.
  • Potential for Error Propagation: The performance of the W-Net can be limited by the accuracy of the first V-Net's skeleton mask. Improvements in the initial segmentation could lead to better overall performance.

The authors are optimistic that with increased data availability and further development, deep learning methods will play a significant role in automated MM lesion detection for population studies and clinical practice.

Summary updated

Abstract

The identification of bone lesions is crucial in the diagnostic assessment of multiple myeloma (MM). <sup>68</sup>Ga-Pentixafor PET/CT can capture the abnormal molecular expression of CXCR-4 in addition to anatomical changes. However, whole-body detection of dozens of lesions on hybrid imaging is tedious and error prone. It is even more difficult to identify lesions with a large heterogeneity. This study employed deep learning methods to automatically combine characteristics of PET and CT for whole-body MM bone lesion detection in a 3D manner. Two convolutional neural networks (CNNs), V-Net and W-Net, were adopted to segment and detect the lesions. The feasibility of deep learning for lesion detection on <sup>68</sup>Ga-Pentixafor PET/CT was first verified on digital phantoms generated using realistic PET simulation methods. Then the proposed methods were evaluated on real <sup>68</sup>Ga-Pentixafor PET/CT scans of MM patients. The preliminary results showed that deep learning method can leverage multimodal information for spatial feature representation, and W-Net obtained the best result for segmentation and lesion detection. It also outperformed traditional machine learning methods such as random forest classifier (RF), <i>k</i>-Nearest Neighbors (<i>k</i>-NN), and support vector machine (SVM). The proof-of-concept study encourages further development of deep learning approach for MM lesion detection in population study.
BibTeX
@inproceedings{xu2018automated,
  author = {Lina Xu and Giles Tetteh and Jana Lipková and Yu Zhao and Hongwei Li and Patrick Ferdinand Christ and Marie Piraud and Andreas K. Buck and Kuangyu Shi and Bjoern Menze},
  title = {{Automated Whole-Body Bone Lesion Detection for Multiple Myeloma on <sup>68</sup>Ga-Pentixafor PET/CT Imaging Using Deep Learning Methods}},
  booktitle = {Contrast},
  year = {2018},
  doi = {10.1155/2018/2391925},
}
Figure