School of Computer and Information Engineering, Central South University of Forestry and Technology, Changsha 410082, Hunan, China.
College of Information Engineering, Changsha Medical University, Changsha 410219, Hunan, China.
Comput Intell Neurosci. 2022 Aug 4;2022:5863782. doi: 10.1155/2022/5863782. eCollection 2022.
The object detection task in the medical field is challenging in terms of classification and regression. Due to its crucial applications in computer-aided diagnosis and computer-aided detection techniques, an increasing number of researchers are transferring the object detection techniques to the medical field. However, in existing work on object detection, researchers do not consider the low resolution of medical images, the high amount of noise, and the small size of the objects to be detected. Based on this, this paper proposes a new algorithmic model called the MS Transformer, where a self-supervised learning approach is used to perform a random mask on the input image to reconstruct the input features, learn a richer feature vector, and filter out excessive noise. To focus the model on the small objects that are being detected, the hierarchical transformer model is introduced in this paper, and a sliding window with a local self-attention mechanism is used to give a higher attention score to the small objects to be detected. Finally, a single-stage object detection framework is used to predict the sequence of sets at the location of the bounding box and the class of objects to be detected. On the DeepLesion and BCDD benchmark dataset, the model proposed in this paper achieves better performance improvement on multiple evaluation metric categories.
医学领域中的目标检测任务在分类和回归方面具有挑战性。由于其在计算机辅助诊断和计算机辅助检测技术中的关键应用，越来越多的研究人员将目标检测技术应用于医学领域。然而，在现有的目标检测工作中，研究人员没有考虑到医学图像的低分辨率、高噪声和待检测目标的小尺寸等问题。基于此，本文提出了一种名为 MS Transformer 的新算法模型，该模型采用自监督学习方法对输入图像进行随机遮挡，以重建输入特征，学习更丰富的特征向量，并滤除过多的噪声。为了使模型专注于要检测的小目标，本文引入了分层 Transformer 模型，并使用具有局部自注意力机制的滑动窗口为要检测的小目标赋予更高的注意力得分。最后，使用单阶段目标检测框架预测边界框位置处的集合序列和要检测的对象类别。在 DeepLesion 和 BCDD 基准数据集上，本文提出的模型在多个评估指标类别上实现了更好的性能提升。
