Abstract: Manual detection of spinal lesions is a time-consuming task that heavily relies on the expertise of specialists in the field. Consequently, the automatic identification of spinal lesions has become essential. Developing accurate automated systems for the detection and classification of spinal lesions is therefore crucial. However, this endeavor presents significant challenges due to considerable variability in lesion size, location, and structure. Additionally, spinal tumors often exhibit high radiological similarity to the rare disease Brucellosis, which can further complicate diagnosis. To address these challenges, this paper proposes a novel and enhanced spinal lesions MRI images recognition model. Firstly, we introduce a bi-directional feature pyramid network backbone based on resnet-101. Deformable convolution is utilized instead of traditional convolution across various layers to extract richer semantic features. Concurrently, multiple attention mechanisms including self-attention mechanisms and soft attention mechanisms are integrated into different modules to effectively fuse the most informative feature components. Finally, an improved balanced cross-entropy loss function is implemented to alleviate the data imbalance issue arising from the rarity of both spinal tumors and Brucellosis. Validation conducted on a clinical dataset provided by the Second Affiliated Hospital of Dalian Medical University demonstrates that our recognition precision rate reached 94.2%, while the recall rate achieved 90.8%. Experimental results indicate that the proposed method outperforms other models in terms of recognition performance.

Key words: spinal lesions recognition, bi-directional feature pyramid network, multi-attention mechanisms, deformable convolution, multi-feature fusion

摘要： 人工检测脊柱病变是一项耗时的工作，并且高度依赖于该领域的专家，因此，脊柱病灶的自动识别是非常必要的。然而，脊柱病灶的准确定位和分类是一项具有挑战性的工作，因为脊柱病灶的大小、位置和结构存在着广泛的差异，同时脊柱肿瘤与稀有病布鲁氏菌在影像上高度相似。为了应对这些挑战，本文提出了一种改进的脊柱病灶MRI影像识别模型，首先引入以resnet-101为基础的双向特征金字塔主干网络，并利用可变卷积在不同层替代传统的卷积神经网络，能从特征层中获得更多的特征信息。同时在不同的模块中加入了多重注意力，包括自注意力机制和柔性注意力机制，有效的融合特征中贡献较大的部分，最后为了克服脊柱肿瘤、感染性病变、稀有病布鲁氏菌的数据不平衡问题，引入了改进的平衡交叉熵损失函数。在大连医科大学第二附属医院提供的临床数据集上进行验证，识别精确率达到了94.2%，识别召回率达到90.8%。与其他识别模型进行对比实验，结果也说明本文方法相对于其他模型识别性能更好。

关键词: 脊柱病灶识别, 双向特征金字塔, 多注意力机制, 可变卷积, 多特征融合