Annu Int Conf IEEE Eng Med Biol Soc. 2021 Nov;2021:3374-3377. doi: 10.1109/EMBC46164.2021.9630988.
In the last decades, MRI was proven a useful tool for the diagnosis and characterization of Prostate Cancer (PCa). In the literature, many studies focused on characterizing PCa aggressiveness, but a few have distinguished between low-aggressive (Gleason Grade Group (GG) <=2) and high-aggressive (GG>=3) PCas based on biparametric MRI (bpMRI). In this study, 108 PCas were collected from two different centers and were divided into training, testing, and validation set. From Apparent Diffusion Coefficient (ADC) map and T2-Weighted Image (T2WI), we extracted texture features, both 3D and 2D, and we implemented three different methods of Feature Selection (FS): Minimum Redundance Maximum Relevance (MRMR), Affinity Propagation (AP), and Genetic Algorithm (GA). From the resulting subsets of predictors, we trained Support Vector Machine (SVM), Decision Tree, and Ensemble Learning classifiers on the training set, and we evaluated their prediction ability on the testing set. Then, for each FS method, we chose the best classifier, based on both training and testing performances, and we further assessed their generalization capability on the validation set. Between the three best models, a Decision Tree was trained using only two features extracted from the ADC map and selected by MRMR, achieving, on the validation set, an Area Under the ROC (AUC) equal to 81%, with sensitivity and specificity of 77% and 93%, respectively.Clinical Relevance- Our best model demonstrated to be able to distinguish low-aggressive from high-aggressive PCas with high accuracy. Potentially, this approach could help clinician to noninvasively distinguish between PCas that might need active treatment and those that could potentially benefit from active surveillance, avoiding biopsy-related complications.
在过去的几十年中，MRI 已被证明是诊断和表征前列腺癌（PCa）的有用工具。在文献中，许多研究都集中在表征 PCa 的侵袭性上，但只有少数研究根据双参数 MRI（bpMRI）区分低侵袭性（Gleason 分级组（GG）<=2）和高侵袭性（GG>=3）PCa。在这项研究中，从两个不同的中心收集了 108 个 PCa，并将其分为训练集、测试集和验证集。从表观扩散系数（ADC）图和 T2 加权图像（T2WI）中，我们提取了 3D 和 2D 纹理特征，并实现了三种不同的特征选择（FS）方法：最小冗余最大相关性（MRMR）、亲和传播（AP）和遗传算法（GA）。从得到的预测因子子集中，我们在训练集上使用支持向量机（SVM）、决策树和集成学习分类器对其进行训练，并在测试集上评估其预测能力。然后，对于每种 FS 方法，我们根据训练和测试性能选择最佳分类器，并进一步在验证集上评估其泛化能力。在这三个最佳模型中，使用仅从 ADC 图中提取的两个特征并由 MRMR 选择的决策树进行训练，在验证集上达到了 81%的 AUC，敏感性和特异性分别为 77%和 93%。临床相关性-我们的最佳模型证明能够以高精度区分低侵袭性和高侵袭性 PCa。潜在地，这种方法可以帮助临床医生无创地区分可能需要积极治疗的 PCa 和那些可能从主动监测中受益的 PCa，避免活检相关的并发症。
