The RADLER framework is demonstrated with a DL ar- chitecture for CT/PET; in detail, a 3D multimodal CNN is adapted from a 2D solution originally aimed at classifying lung nodules from CT imaging [39]. Secondly, we adopted an internal transfer learning approach, starting from the diag- nostic classification of tumour stage. This domain adaptation approach proved useful in dealing with class unbalance and a relatively low number of samples, while achieving good predictive performance, as shown on the HN dataset, with high class unbalance and low number of samples. 解释

这段话大概是在说明RADLER框架采用了一个针对CT/PET的DL架构；具体而言，是从最初旨在从CT成像中对肺结节进行分类的2D解决方案中改编的3D多模态CNN。其次，我们采用了一种内部迁移学习方法，从肿瘤分期的诊断分类开始。这种领域自适应方法在处理类别不平衡和相对较少的样本时证明是有用的，同时在HN数据集上取得了良好的预测性能，该数据集具有高度类别不平衡和少量样本。

相关问题

The RADLER framework introduced in this study aims at the integration of deep and radiomics features for medical image analysis and classification. Its first application in a prognostic task of locoregional recurrence (LR task) of head and neck cancer improves with respect to the state of art [38], both in terms of sensitivity (0.94 vs 0.56) and specificity (0.95 vs 0.67). Moreover, the DAP included in the framework is used to evaluate variability due to resampling and control for selection bias in the model selection phase. As assessed by the DAP , the feature set integrating radiomics and deep features is more effective in predicting LR than only one of the feature types. 解释

这段话大概是在说明一个研究的框架RADLER旨在将深度学习和放射学特征集成到医学图像分析和分类中。其首个应用于头颈癌局部复发（LR任务）的预后任务中，在灵敏度（0.94 vs 0.56）和特异度（0.95 vs 0.67）方面都比现有技术有所提高。此外，框架中包含的DAP用于评估由于重采样引起的变异性，并控制模型选择阶段中的选择偏差。通过DAP评估，将放射学和深度特征集成到特征集中比仅使用其中一种特征类型更有效地预测LR。