深入医学图像分析综述

深入医学图像分析

Altaf F., Islam S.M.S., Akhtar N., et al. Delving into advanced medical imaging techniques: exploring concepts and methods while addressing challenges and outlining future directions[J]. IEEE Access, 2019.

医学图像分析的深度学习方法

检测/定位，分割，配准，分类

深入医学图像分析的挑战

• 数据缺乏充分标注的情况下，在有限数据集上构建复杂模型可能导致其对训练数据的高度敏感（overfitting）。
• 样本分布失衡的问题会导致分类器在某些类别上出现性能下降。
• 无法提供完整的预测区间是深度学习文献中的常见问题。通常情况下，在处理分类任务时，默认的结果被视为概率估计。
  在深度学习领域中, 模型的输出结果通常被解释为概率估计值, 这使得结果解读存在局限性. 尤其是在涉及高风险决策的任务中, 缺少预测值范围信息会带来严重后果.

未来的方向

By organizing diverse tasks that enable the deep models to transfer effectively among each other, they have developed a referred-to-as-task taxonomy aimed at providing guidance on applying transfer learning techniques specifically for natural image processing.

Using existing deep models as tools for feature extraction and further developing these features presents a much more promising approach without relying on bigger datasets.

The process of pointing towards a direction requires post-processing of deep features to better meet the demands of the underlying Medical Image Analysis task.

Training Partially Frozen Deep Networks

Using GANs for Synthetic Data Generation

Miscellaneous Data Augmentation Techniques