Automatic Localization of the Subthalamic Nucleus on Patient-Specific Clinical MRI by Incorporating 7T MRI and Machine Learning: Application in Deep Brain Stimulation

Jinyoung Kim, Yuval Duchin, Reuben R. Shamir, Rémi Patriat, Jerrold L. Vitek, Noam Harel, Guillermo SapirobioRxiv 2018 · DOI

TL;DR

This paper presents a novel machine learning framework (7T-ML) that leverages high-field 7T MRI data to accurately localize the subthalamic nucleus (STN) on standard, lower-field clinical MRI scans. This approach aims to improve the precision of Deep Brain Stimulation (DBS) targeting for Parkinson's disease without requiring expensive 7T scanners in every clinic.

What Problem Does This Paper Address?

Deep Brain Stimulation (DBS) is a crucial treatment for Parkinson's disease, but its effectiveness hinges on precisely targeting the subthalamic nucleus (STN). Standard clinical MRI scans (1.5T or 3T) often lack the resolution to clearly visualize the STN and its surrounding structures, making accurate targeting difficult. While intraoperative microelectrode recordings (MER) can improve accuracy, they are time-consuming and require specialized expertise. High-field 7T MRI offers superior visualization but is not widely available in clinical settings. This paper addresses the challenge of achieving accurate, patient-specific STN localization on readily available clinical MRI by bridging the gap between high-quality 7T data and routine clinical practice.

Methods

The core of the proposed method, termed "7T-ML," involves training a machine learning model on a dataset of paired 7T and clinical MRI scans. The process can be broken down into these key steps:

  1. 7T Data Modeling: Using a large database of 7T MRI scans and their corresponding clinical MRI pairs, the researchers first model the spatial and geometric relationships between the STN and its adjacent structures (termed "predictors"). This is achieved using machine learning algorithms that capture these dependencies.
  2. Predictor Segmentation: For a new patient's clinical MRI, the algorithm automatically identifies and segments these "predictor" structures.
  3. STN Prediction: The learned geometric relationships from the 7T data are then applied to the segmented predictors on the clinical MRI to predict the patient-specific STN location and shape.
  4. Ensemble and Robust Prediction: The method utilizes an ensemble approach, combining predictions from multiple training subsets. A "robust prediction" strategy further refines this by weighting training subsets based on their estimated contribution to the prediction accuracy.

The paper also details preprocessing steps like co-registration of 7T and clinical MRIs and the use of advanced normalization tools for image registration.

Datasets

The study utilized a dataset of 80 subjects, each with paired 7T MRI and clinical MRI (1.5T or 3T) scans.

  • Training Set: 46 pairs of 7T and 1.5T MRIs from patients were used for training.
  • Validation Set: 34 pairs of 7T and 1.5T or 3T MRIs from patients and normal controls were used for validation.
  • In-depth Study: A subset of 15 PD patients from the training set was used for detailed analysis of factors affecting prediction.

The authors also compared their method against several publicly available STN atlases: 7T3T, UHFA, DISTAL, and MNIPD25.

Key Insights and Ideas

  • Bridging the 7T-Clinical Gap: The most significant insight is the successful transfer of anatomical knowledge from high-resolution 7T MRI to standard clinical MRI. This bypasses the need for 7T scanners in every clinical setting.
  • Predictors are Key: The study highlights the importance of surrounding structures ("predictors") in accurately localizing the STN, even when the STN itself is not clearly visible on clinical scans.
  • Robustness of the Method: The 7T-ML framework demonstrates robustness, providing comparable accuracy across different clinical MRI field strengths (1.5T and 3T), even though it was not trained on 3T data.
  • Clinical Relevance of Sub-region Accuracy: The paper emphasizes that accurate sub-region STN localization is crucial for optimizing DBS targeting and predicting active contact locations, which directly impacts clinical outcomes.

Key Results

The 7T-ML method demonstrated superior performance compared to state-of-the-art atlas-based methods:

  • Center of Mass Distance: The 7T-ML achieved an average center of mass distance of 1.25±0.60mm, significantly closer to the 7T manual ground truth than other atlases. 89.4% of STNs predicted by 7T-ML were within 2mm of the ground truth, a clinically acceptable threshold.
  • Surface Point Distance and Dice Coefficient (DC): 7T-ML also yielded significantly better mean surface point distances (0.57±0.18mm) and DC values (0.64±0.12%) compared to atlases.
  • Volume: The predicted STN volumes from 7T-ML were comparable to the 7T manual ground truth and the 7T3T atlas.
  • Active Contact Localization: The 7T-ML accurately predicted active contact locations, particularly in the dorsal-lateral region, which is a critical motor territory for DBS. This accuracy was superior to atlas-based methods, which often showed large errors in shape and location, potentially leading to misleading electrode placement.

Limitations and Future Work

  • Geometric Distortion: The authors acknowledge that geometric distortion in 7T MRI and inaccurate co-registration between 7T and clinical data can affect the quality of the "7T priors" and introduce errors. Minimizing these biases was a critical focus.
  • STN Boundary Visibility: The paper notes that the STN's ventral border, adjacent to the substantia nigra (SN), can be challenging to identify automatically on standard clinical MRI due to consistent intensity variations. Future work aims to refine STN shape prediction by automatically identifying this boundary.
  • Generalizability: While the study used a diverse dataset, further investigation into the consistency of STN boundary intensities across a larger population and different centers is warranted.
  • Broader Applications: The authors suggest that the presented framework could be extended to segment other important brain structures for DBS, such as the internal globus pallidus and Vim.
  • Deep Learning Integration: The potential to apply deep learning architectures for segmenting predictor structures is mentioned as a future direction.

Summary updated

Abstract

ABSTRACT Deep Brain Stimulation (DBS) of the subthalamic nucleus (STN) has shown clinical potential for relieving the motor symptoms of advanced Parkinson’s disease. While accurate localization of the STN is critical for consistent across-patients effective DBS, clear visualization of the STN under standard clinical MR protocols is still challenging. Therefore, intraoperative microelectrode recordings (MER) are incorporated to accurately localize the STN. However, MER require significant neurosurgical expertise and lengthen the surgery time. Recent advances in 7T MR technology facilitate the ability to clearly visualize the STN. The vast majority of centers, however, still do not have 7T MRI systems, and fewer have the ability to collect and analyze the data. This work introduces an automatic STN localization framework based on standard clinical MRIs without additional cost in the current DBS planning protocol. Our approach benefits from a large database of 7T MRI and its clinical MRI pairs. We first model in the 7T database, using efficient machine learning algorithms, the spatial and geometric dependency between the STN and its adjacent structures (predictors). Given a standard clinical MRI, our method automatically computes the predictors and uses the learned information to predict the patient-specific STN. We validate our proposed method on clinical T 2 W MRI of 80 subjects, comparing with experts-segmented STNs from the corresponding 7T MRI pairs. The experimental results show that our framework provides more accurate and robust patient-specific STN localization than using state-of-the-art atlases. We also demonstrate the clinical feasibility of the proposed technique assessing the post-operative electrode active contact locations.
BibTeX
@misc{kim2018automatic,
  author = {Jinyoung Kim and Yuval Duchin and Reuben R. Shamir and Rémi Patriat and Jerrold L. Vitek and Noam Harel and Guillermo Sapiro},
  title = {{Automatic Localization of the Subthalamic Nucleus on Patient-Specific Clinical MRI by Incorporating 7T MRI and Machine Learning: Application in Deep Brain Stimulation}},
  year = {2018},
  doi = {10.1101/322230},
  howpublished = {bioRxiv},
}

Loading related content...

Figure