文献速递：深度学习--应用深度学习到DaTscan SPECT图像以改善帕金森病运动结果预测

文献速递：基于Deep Learning的PaCDaTS图像分析--一种优化帕金森病运动结果预测的方法

Title

题目

Enhanced accuracy in predicting motor outcomes for Parkinson’s disease patients using advanced deep learning techniques in DaTscan SPECT imaging studies.

应用深度学习到DaTscan SPECT图像以改善帕金森病运动结果预测

01

文献速递介绍

parkin森病（PD）是神经退行性疾病的一种常见类型。当病情进一步发展时可能会导致一系列症状如运动迟缓静止性震颤僵硬以及姿态不稳等综合表现进而导致运动功能障碍。尽管 parkin森确切病因尚不清楚但它与黑质内多巴胺能神经元减少相关目前针对该病症缺乏有效的治疗方法因此寻找可靠的 parkin森疾病进展生物标志物具有重要意义。由于 parkin森患者的症状表现形式多样且病情进展速度不一如何预测疾病进展有助于优化临床试验方案因此相关研究也备受关注。

多巴胺转运体（DAT）SPECT成像（临床常被称为DaTscan）已被广泛应用于帕金森病（PD）的诊断工作中，并为早期患病个体提供了重要的参考价值；这类患者往往表现出不明确的帕金森病综合症症状表现。通过图像获得准确度适宜的结果后，则有助于通过比较实际结果与预测结果来评估治疗效果的时间可行性更加便捷

DAT SPECT图像在临床应用中以视觉检查为核心环节，并通过定量分析辅助解决早期检测中的挑战及推进疾病监测与随访工作。基于放射组学的研究不仅显著提升了诊断效能、预后评估以及治疗反应评估，在推进疾病监测与随访方面也取得了显著成效。同样展现了类似的优势。
研究开发了结合影像学特性和非影像学特征的数据融合模型，在预测患者的运动功能表现及相关认知下降风险方面取得了突破性进展。
我们的前期研究重点在于利用DAT SPECT图像的放射组学特性及其非成像临床指标来构建预测模型，并成功识别出对预测性能具有决定性影响的关键指标。
这些探索不仅带来了理论上的新思路，并为实际应用提供了可行性支持。

Methods

方法

2.1. Patient data

Long-term data were obtained from the PPMI database, which is accessible at www.ppmi-info.org/data. The Parkinson Progression Marker Initiative in 2011 provided the necessary information.

Utilizing the UPDRS – part III (motor) score at year 4, DAT_SPECT imaging data along with UPDRS_III scores from baseline and subsequent years were utilized to predict motor symptoms.

1. These measures were chosen as inputs because our earlier studies showed that they were the most relevant to prediction performance. Other clinical features such as gender, age, MoCA, and disease duration at times of diagnosis and appearance of symptoms were irrelevant . A UPDRS_III score was only considered valid if the patient was not on medication at the time of evaluation. Any patient that had all of the aforementioned data available from the PPMI database was included in this study, resulting in 252 PD subjects (baseline age of 62.4 ± 9.7 be tween 35 and 86; 168 male, 84 female) with year 4 outcome UPDRS_III distribution of 28.7 ± 11.7, ranging between 5 and 77. All images collected in the PPMI database followed a standard acquisition protocol, to account for different SPECT systems used throughout the sites participating in the study. Imaging was done 4 ± 0.5 h after a dosage of 111–185 MBq radiotracer injection. Thyroid protection was provided by pre-treating subjects with an iodine solution before radiotracer injection. The data, in the form of a 128 × 128 matrix, was acquired over a total scan duration of about 30–45 min by sampling every 3◦, 120 projections, 20% symmetric photopeak windows centered on 159 keV and 122 keV. The raw SPECT projection data was recon structed using the iterative OSEM (ordered-subset expectation maximi zation) method on a HERMES system (Hermes Medical Solutions, Stockholm, Sweden). PMOD (PMOD Technologies, Zurich, Switzerland) was then used to apply Chang’s attenuation correction to the recon structed data. PMOD was also used for spatial normalization into Montreal Neurological Institute (MNI) space based on a multisite Eu ropean database of healthy control patients [33]. This resulted in a size of 91 × 109 × 91 voxels for the final images, with voxel size of 2 × 2 × 2 mm3 . As an example, DAT SPECT images of two subjects (a male and a female) taken at years 0 and 1 are shown in Fig. 1. Patient (a) saw his UPDRS_III score progress from 41 to 48 from baseline to year 4 while patient (b)’s score went from 13 to 22 over the same duration. There seem to be image characteristics that could be related to the scores. We expect the method developed in this work to make use of these characteristics.

2.1. 患者数据

18, 21

18, 21

33

引用

Results

结果

Using only UPDRS_III scores from year 0 and year 1, the predicted motor scores at year 4 exhibited an average discrepancy of 7.6 ±_ _6.1 compared to actual values (range ). Incorporating longitudinal DAT SPECT images reduced this average discrepancy to 6.0 ± 4.8, which proved to be statistically significantly improved through a two-sample t-test, thereby highlighting the value of including such imaging data.

基于第0年和第1年的UPDRS_III评分类别进行预测研究发现其与真实第四年运动评分之间的平均偏差为7.6±6.1（范围）。引入随访时期的DAT SPECT图像分析后其平均偏差降至6.0±4.8经两组样本t检验分析可知该差异具有统计学显著性意义提示加入影像数据有助于提高预测精度。

Conclusion

结论

This research demonstrates that incorporating DAT SPECT images alongside UPDRS_III scores into deep learning-based prediction models significantly enhances their performance. By employing a CNN-based prediction approach, simpler and more universal applications are enabled without the need for segmentation or feature extraction.

研究表明将DAT SPECT图像加入到UPDRS_III评分中被用作深度学习预测模型的输入数据显著改善了结果。基于CNN的方法无需对图像进行分割或特征提取即可实现更为简便且具有广泛适用性的预测。

Figure

图

Figure 1 shows Examples of longitudinal DAT SPECT images from two patients (i, ii), acquired at year 0 (1) and year 1 (2).

图 1. 两名患者（a，b）在第0年（1）和第1年（2）收集的纵向DAT SPECT图像示例。

The CNN architecture was employed to predict the UPDRS_III scores at the fourth year, derived from longitudinal DAT SPECT images and UPDRS_III ratings.

该研究基于纵向DAT SPECT图像和UPDRS-III评分这一指标体系，并旨在预测第四年的UPDRS-III评分。

The process of training and testing aimed to predict UPDRS_III at year 4 based on longitudinal DAT SPECT images and UPDRS_III scores.

图 3. 基于纵向方向上的DAT SPECT影像及UPDRS-III评分预测第四年UPDRS-III评分的训练与测试流程

Fig. 4 illustrates the comparison between forecasted and actual UPDRS_III outcomes at year 4, specifically when relying solely on UPDRS_III scores (a) and when incorporating both UPDRS_III and DAT SPECT image data (b). The dataset incorporates information from both year 0 and year 1.

图 4展示了两种不同的测试方案及其效果比较：在单一条件下采用单组测试方案（a），以及结合单组测试方案与图像评估方法（b）。研究通过第四年的预测结果与实际值进行了对比分析，并包含基线数据（第0年）以及一年后的随访数据（第1年）。

The outcomes at year 4 were analyzed in two scenarios: one utilizing only the UPDRS_III scores from the baseline year (labeled as case a), and another incorporating both the UPDRS_III scores and the DAT SPECT images from the same baseline year (labeled as case b).

图5展示了仅基于第0年UPDRS-III评分为输入（a），以及同时考虑了第0年UPDRS-III得分与DAT SPECT图像结合（b）的情况。通过4年的预测评估与实际值的比较

The predicted versus actual outcomes of UPDRS III were evaluated at four years after implementing the model, with two distinct scenarios being analyzed: one that relied solely on the year 1 scores of UPDRS III and another that incorporated both the UPDRS III and DATSPECT images from the first year.

图 6展示了两种不同的实验设置对比：仅采用第一年 UPDRSⅢ 得分进行预测 (a)，以及同时结合第一年 UPDRSⅢ 得分以及 DAT/SPECT 图像的数据进行分析 (b)，并比较这两种方法在第四年的预测结果与实际值的匹配程度

Figure 7 illustrates the convergence behavior of a CNN trained on two distinct sets: the training set labeled as class A and the test set labeled as class B when compared across two distinct scenarios: one where the DAT SPECT images are excluded... The use of different colors corresponds to individual cross-validation folds, achieving an average error rate.

both those lacking (1) and those incorporating (2) DAT SPECT images. Each color corresponds to a cross-validation fold, and the loss is quantified through mean average error.*

图中展示的是用于预测案例的CNN收敛情况,涵盖未采用(DAT SPECT无标记数据)与采用(DAT SPECT标签数据)的情况进行比较,将训练数据集分为两部分:未使用的(DAT SPECT无标记数据)作为验证集(a),而被用来训练模型的数据则来自(DAT SPECT标签数据)(b).每一种颜色代表一次交叉验证循环,其中损失值基于平均平均误差计算得出