近年来，随着卷积神经网络的快速发展，基于深度学习的医学影像分析成为了计算机领域研究的热点。然而，由于保护患者隐私、标注成本高及病症样本不足等问题，制作大规模医学图像数据集是非常困难的，深度学习模型无法进行充分训练，从而导致特征提取能力不足。现有的方法主要通过数据增强来增加训练样本，这种做法不但费时费力，还可能引入噪声数据。此外，医学图像具有复杂的结构特征和较高的类间相似度，使得分类难度进一步增加。因此，如何利用有限标注样本对医学图像进行准确分类成为了当前的研究热点之一。

（1）针对医学图像标注数据较少导致模型特征提取能力不足的问题，本文提出了一种基于空间变换网络的小样本医学图像分类模型STW-ResNet。首先，模型采用了“预训练微调+元学习”的小样本学习架构，并设计了训练难度递进的两阶段预训练策略，使模型能够更有效地学习源域特征。其次引入空间变换网络模块，通过仿射变换自适应放大病变区域，增强网络对图像空间变化的适应能力，提高边缘特征提取能力。最后，提出结合特征分布校准和最邻近质心算法的特征变换分类器，简化分类过程的同时提升了分类精度。在ISIC2018皮肤病变数据集和Pap-smear宫颈细胞数据集上对模型进行验证，与基线模型相比，STW-ResNet在ISIC2018数据集上的平均分类精度提高了2.50%；在Pap-smear数据集上的平均分类精度提高了3.91%。

（2）针对医学图像复杂度高、类间相似度大导致模型分类精度不高的问题，本文提出了一种基于多尺度小波特征融合的双分支小样本医学图像分类模型MSTWs-ResNet。首先，通过引入多尺度特征提取网络，捕获图像在不同尺度下的特征信息，有效提高了模型对于复杂特征的识别能力，并结合小波变换特征融合模块，解决不同尺度特征融合时产生的信息丢失问题；其次，将多尺度特征提取网络与改进WideResNet结合，构建双分支特征提取网络来进一步增强模型的特征提取能力；最后，引入稀疏轴向MLP，降低模型参数量的同时，高效建立图像的全局依赖关系，增强对相似特征的区分能力。实验结果表明，MSTWs-ResNet能够在复杂特征和相似特征上提取到有效特征，在两个数据集上都取得了较好的分类精度。与基线模型相比，在ISIC2018数据集上的平均分类精度提高了3.20%；在Pap-smear数据集上的平均分类精度提高了6.73%。

In recent years, with the rapid development of convolutional neural networks, medical image analysis based on deep learning has become a hot topic in computer science research. However, due to issues such as protecting patient privacy, high annotation costs, and insufficient disease samples, it is very difficult to produce large-scale medical image datasets, and deep learning models cannot be fully trained, resulting in insufficient feature extraction capabilities. Existing methods mainly increase training samples through data augmentation, which is not only time-consuming and labor-intensive, but may also introduce noisy data. In addition, medical images have complex structural features and high inter-class similarity, which further increases the difficulty of classification. Therefore, how to accurately classify medical images using limited annotated samples has become one of the current research hotspots.

(1) In response to the problem of insufficient model feature extraction capabilities due to the lack of annotated data for medical images, this paper proposes a few shot medical image classification model STW-ResNet based on a spatial transformer network. First, the model adopts a few shot learning architecture of "pre-training fine-tuning + meta-learning" and designs a two-stage pre-training strategy with progressive training difficulty, so that the model can learn source domain features more effectively. Secondly, the spatial transformer network module is introduced to adaptively enlarge the lesion area through affine transformation, enhance the network's adaptability to image spatial changes, and improve the edge feature extraction capability. Finally, a feature transformation classifier combining feature distribution calibration and nearest centroid algorithm is proposed to simplify the classification process and improve the classification accuracy. The model is verified on the ISIC2018 skin lesion dataset and the Pap-smear cervical cell dataset. Compared with the baseline model, the average classification accuracy of STW-ResNet on the ISIC2018 dataset is improved by 2.50%; the average classification accuracy on the Pap-smear dataset is improved by 3.91%.

(2) In order to solve the problem of low model classification accuracy due to high complexity of medical images and large similarity between classes, this paper proposes a dual-branch few shot medical image classification model MSTWs-ResNet based on multi-scale wavelet feature fusion. Firstly, by introducing a multi-scale feature extraction network, the feature information of the image at different scales is captured, which effectively improves the model's ability to recognize complex features. In addition, the wavelet transform feature fusion module is combined to solve the problem of information loss when different scale features are fused. Secondly, the multi-scale feature extraction network is combined with the improved WideResNet to construct a dual-branch feature extraction network to further enhance the model's feature extraction ability. Finally, the sparse axial MLP is introduced to reduce the number of model parameters while efficiently establishing the global dependency of the image and enhancing the ability to distinguish similar features. Experimental results show that MSTWs-ResNet can extract effective features from complex features and similar features, and achieves good classification accuracy on both datasets. Compared with the baseline model, the average classification accuracy on the ISIC2018 dataset is improved by 3.20%; the average classification accuracy on the Pap-smear dataset is improved by 6.73%.

[1] 黄一超,傅锐芝,王昕辰,等.医学影像图像与组织学图像配准的研究进展[J].中国医学计算机成像杂志,2024,30(05):646-652.DOI:10.19627/j.cnki.cn31-1700/th.2024.05.014.

[2] 赵愉,王得旭,顾力栩.人工智能技术在计算机辅助诊断领域的发展新趋势[J].中国科学:生命科学,2020,50(11):1321-1334.

[3] Alonso-Martínez J L, Sánchez F J A, Echezarreta M A U. Delay and misdiagnosis in sub-massive and non-massive acute pulmonary embolism[J]. European journal of internal medicine, 2010, 21(4): 278-282.

[4] Hendriksen J M T, Koster-van Ree M, Morgenstern M J, et al. Clinical characteristics associated with diagnostic delay of pulmonary embolism in primary care: a retrospective observational study[J]. BMJ open, 2017, 7(3): e012789.

[5] Li Z, Tang H, Peng Z, et al. Knowledge-Guided Semantic Transfer Network for Few-Shot Image Recognition[J]. IEEE Transactions on Neural Networks and Learning Systems, 2023.

[6] Diwan T, Anirudh G, Tembhurne J V. Object detection using YOLO: Challenges, architectural successors, datasets and applications[J]. Multimedia Tools and Applications, 2022: 1-33.

[7] Cheng B, Misra I, Schwing AG, et al. Masked-attention mask transformer for universal image segmentation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022: 1290-1299.

[8] Esteva A, Kuprel B, Novoa R A, et al. Dermatologist-level classification of skin cancer with deep neural networks[J]. nature, 2017, 542(7639): 115-118.

[9] Rakocz N, Chiang J N, Nittala M G, et al. Automated identification of clinical features from sparsely annotated 3-dimensional medical imaging[J]. NPJ digital medicine, 2021, 4(1): 44.

[10] 陈鹏.中国医疗人工智能现状分析：从产品验证进入市场验证[J].互联网经济,2020,(Z1):86-91.DOI:10.19609/j.cnki.cn10-1255/f.2020.z1.016.

[11] Rahman M M, Davis D N. Addressing the class imbalance problem in medical datasets[J].International Journal of Machine Learning and Computing, 2013, 3(2):224.

[12] Li D C, Liu C W, Hu S C. A learning method for the class imbalance problem with medical data sets[J]. Computers in biology and medicine, 2010, 40(5): 509-518.

[13] Chen R, Huang J, Song Y, et al. Deep learning algorithms for brain disease detection with magnetic induction tomography[J]. Medical Physics, 2021, 48(2): 745-759.

[14] Abiwinanda N, Hanif M, Hesaputra S T, et al. Brain tumor classification using convolutional neural network[C]//World Congress on Medical Physics and Biomedical Engineering 2018: June 3-8, 2018, Prague, Czech Republic (Vol. 1). Springer Singapore, 2019: 183-189.

[15] Afshar P, Plataniotis K N, Mohammadi A. Capsule networks for brain tumor classification based on MRI images and coarse tumor boundaries[C]//ICASSP 2019-2019 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2019: 1368-1372.

[16] Liu M, Zhang J, Lian C, et al. Weakly supervised deep learning for brain disease prognosis using MRI and incomplete clinical scores[J]. IEEE transactions on cybernetics, 2019, 50(7): 3381-3392.

[17] Xu W, Xu Y, Chang T, et al. Co-scale conv-attentional image transformers[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021.

[18] Nawaz M, Nazir T, Javed A, et al. An efficient deep learning approach to automatic glaucoma detection using optic disc and optic cup localization[J]. Sensors, 2022, 22(2): 434.

[19] Peng Y, Dharssi S, Chen Q, et al. DeepSeeNet: a deep learning model for automated classification of patient-based age-related macular degeneration severity from color fundus photographs[J]. Ophthalmology, 2019, 126(4): 565-575.

[20] 袁媛,陈明惠,柯舒婷等.基于集成卷积神经网络和 Vit 的眼底图像分类研究[J].中国激光,2022,49(20):108-116.

[21] Qasim Gilani S, Syed T, Umair M, et al. Skin Cancer Classification Using Deep Spiking Neural Network[J]. Journal of Digital Imaging, 2023, 36(3): 1137-1147.

[22] Kaur R, GholamHosseini H, Sinha R, et al. Melanoma classification using a novel deep convolutional neural network with dermoscopic images[J]. Sensors, 2022, 22(3): 1134.

[23] Afza F, Sharif M, Khan M A, et al. Multiclass skin lesion classification using hybrid deep features selection and extreme learning machine[J]. Sensors, 2022, 22(3): 799.

[24] Indraswari R, Rokhana R, Herulambang W. Melanoma image classification based on MobileNetV2 network[J]. Procedia computer science, 2022, 197: 198-207.

[25] Lan Z, Cai S, He X, et al. Fixcaps: An improved capsules network for diagnosis of skin cancer[J]. IEEE Access, 2022, 10: 76261-76267.

[26] Wang B, Zhang W. MARnet: Multi-scale adaptive residual neural network for chest X-ray images recognition of lung diseases[J]. Math. Biosci. Eng, 2022, 19(1): 331-350.

[27] Ali I, Muzammil M, Haq I U, et al. Deep feature selection and decision level fusion for lungs nodule classification[J]. IEEE Access, 2021, 9: 18962-18973.

[28] Fujima N, Andreu-Arasa V C, Onoue K, et al. Utility of deep learning for the diagnosis of otosclerosis on temporal bone CT[J]. European Radiology, 2021, 31: 5206-5211.

[29] Apostolopoulos I D, Aznaouridis S I, Tzani M A. Extracting possibly representative COVID-19 biomarkers from X-ray images with deep learning approach and image data related to pulmonary diseases[J]. Journal of Medical and Biological Engineering, 2020, 40: 462-469.

[30] Marentakis P, Karaiskos P, Kouloulias V, et al. Lung cancer histology classification from CT images based on radiomics and deep learning models[J]. Medical & biological engineering & computing, 2021, 59: 215-226.

[31] Savaş S, Topaloğlu N, Kazcı Ö, et al. Classification of carotid artery intima media thickness ultrasound images with deep learning[J]. Journal of Medical Systems, 2019, 43(8): 273.

[32] Huang G, Liu Z, Van Der Maaten L, et al. Densely connected convolutional networks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 4700-4708.

[33] Howard A G. Mobilenets: Efficient convolutional neural networks for mobile vision applications[J]. arXiv preprint arXiv:1704.04861, 2017.

[34] Cai A, Chen L, Chen Y, et al. Pre-MocoDiagnosis: Few-shot ophthalmic diseases recognition using contrastive learning[C]//2022 IEEE international conference on bioinformatics and biomedicine (BIBM). IEEE, 2022: 2059-2066.

[35] Cano F, Cruz-Roa A. An exploratory study of one-shot learning using Siamese convolutional neural network for histopathology image classification in breast cancer from few data examples[C]//15th international symposium on medical information processing and analysis. SPIE, 2020, 11330: 66-73.

[36] Garg G, Paul A. Few-shot diagnosis of chest x-rays using an ensemble of random discriminative subspaces[J]. arXiv preprint arXiv:2309.00081, 2023.

[37] Guo Z, Wang Y, Liu L, et al. Siamese Network-Based Few-Shot Learning for Classification of Human Peripheral Blood Leukocyte[C]//2021 IEEE 4th International Conference on Electronic Information and Communication Technology (ICEICT). IEEE, 2021: 818-822.

[38] 谢莉,舒卫平,耿俊杰,等.结合加权原型和自适应张量子空间的小样本宫颈细胞分类[J].计算机应用,2024,44(10):3200-3208.

[39] Prabhu V, Kannan A, Ravuri M, et al. Few-shot learning for dermatological disease diagnosis[C]//Machine Learning for Healthcare Conference. PMLR, 2019: 532-552.

[40] Liu X J, Li K, Luan H, et al. Few-shot learning for skin lesion image classification[J]. Multimedia Tools and Applications, 2022, 81(4): 4979-4990.

[41] Finn C, Abbeel P, Levine S. Model-agnostic meta-learning for fast adaptation of deep networks[C]//International conference on machine learning. PMLR, 2017: 1126-1135.

[42] Chao S, Belanger D. Generalizing few-shot classification of whole-genome doubling across cancer types[C]//Proceedings of the IEEE/CVF International Conference on computer vision. 2021: 3382-3392.

[43] Naren T, Zhu Y, Wang M D. COVID-19 diagnosis using model agnostic meta-learning on limited chest X-ray images[C]//Proceedings of the 12th ACM International Conference on Bioinformatics, Computational Biology, and Health Informatics. 2021: 1-9.

[44] 安晨,汪成亮,廖超,等.基于注意力关系网络的无线胶囊内镜图像分类方法[J].计算机工程,2021,47(10):252-259+268.DOI:10.19678/j.issn.1000-3428.0059122.

[45] Liu B, Cao Y, Lin Y, et al. Negative margin matters: Understanding margin in few-shot classification [C]// Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16. Springer International Publishing, 2020: 438-455.

[46] 赵嘉晖.基于高维多目标优化的小样本皮肤癌元学习方法研究[D].太原科技大学,2024.DOI:10.27721/d.cnki.gyzjc.2024.000889.

[47] Mahajan K, Sharma M, Vig L. Meta-dermdiagnosis: Few-shot skin disease identification using meta-learning[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. 2020: 730-731.

[48] Ravi S, Larochelle H. Optimization as a model for few-shot learning[C]//International conference on learning representations. 2017.

[49] Singh R, Bharti V, Purohit V, et al. MetaMed: Few-shot medical image classification using gradient-based meta-learning[J]. Pattern Recognition, 2021, 120: 108111.

[50] Dai Z, Yi J, Yan L, et al. Pfemed: Few-shot medical image classification using prior guided feature enhancement[J]. Pattern Recognition, 2023, 134: 109108.

[51] Hu Y, Gripon V, Pateux S. Leveraging the feature distribution in transfer-based few-shot learning[C]//International Conference on Artificial Neural Networks. Cham: Springer International Publishing, 2021: 487-499.

[52] Chen W Y, Liu Y C, Kira Z, et al. A closer look at few-shot classification[J]. arXiv preprint arXiv:1904.04232, 2019.

[53] 白文鑫.基于少样本学习的医学图像分类算法研究[D].北京邮电大学,2023.DOI:10.26969/d.cnki.gbydu.2023.000122.

[54] Cao J, Yao J, Zhang Z, et al. EFAG-CNN: Effectively fused attention guided convolutional neural network for WCE image classification[C]//2021 IEEE 10th Data Driven Control and Learning Systems Conference (DDCLS). IEEE, 2021: 66-71.

[55] Sunitha S, Sujatha S S. An improved bleeding detection method for wireless capsule endoscopy (wce) images based on alexnet[C]//2021 3rd International conference on signal processing and communication (ICPSC). IEEE, 2021: 11-15.

[56] Wang H, Deng Z H. Cross-domain few-shot classification via adversarial task augmentation[J]. arXiv preprint arXiv:2104.14385, 2021.

[57] Nakamura A, Harada T. Revisiting fine-tuning for few-shot learning[J]. arXiv preprint arXiv:1910.00216, 2019.

[58] Liu H, Tam D, Muqeeth M, et al. Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning[J]. Advances in Neural Information Processing Systems, 2022, 35: 1950-1965.

[59] Maicas G, Bradley A P, Nascimento J C, et al. Pre and post-hoc diagnosis and interpretation of malignancy from breast DCE-MRI[J]. Medical image analysis, 2019, 58: 101562.

[60] Fu W, Chen J, Zhou L. Boosting few-shot rare skin disease classification via self-supervision and distribution calibration[J]. Biomedical Engineering Letters, 2024: 1-13.

[61] Bao H, Dong L, Piao S, et al. Beit: Bert pre-training of image transformers[J]. arXiv preprint arXiv:2106.08254, 2021.

[62] Zheng X, Wang Y, Liu Y, et al. Graph Neural Networks for Graphs with Heterophily: A Survey[J]. arXiv e-prints, 2022: arXiv: 2202.07082.

[63] Zagoruyko S, Komodakis N. Wide Residual Networks[J]. arXiv e-prints, 2016: arXiv: 1605.07146.

[64] Zou J, Ma X, Zhong C, et al. Dermoscopic Image Analysis for ISIC Challenge 2018[J]. arXiv e-prints, 2018: arXiv: 1807.08948.

[65] JANTZEN J, NORUP J, DOUNIAS G, et al. Pap-smear benchmark data for pattern classification[J]. Nature Inspired Smart Information Systems (NiSIS 2005), 2005: 1-9.

[66] Zhang D, Jin M, Cao P. St-metadiagnosis: meta learning with spa tial transform for rare skin disease diagnosis. In 2020 IEEE inter national conference on bioinformatics and biomedicine (BIBM) 2020;2153–2160.

[67] Xing L, Shao S, Liu W, et al. Learning task-specific discriminative embeddings for few-shot image classification[J]. Neurocomputing, 2022, 488: 1-13.

[68] Ma R, Fang P, Drummond T, et al. Adaptive poincaré point to set distance for few-shot classification[C]. Proceedings of the 2022 AAAI Conference on Artificial Intelligence. Palo Alto: AAAI Press, 2022, 36(2): 1926-1934.

[69] 韦世红, 刘红梅, 唐宏, 等. 多级度量网络的小样本学习[J]. 计算机工程与应用,2023,59(02): 94-101.

[70] He K, Zhang X, Ren S, et al. Spatial pyramid pooling in deep convolutional networks for visual recognition[J]. IEEE transactions on pattern analysis and machine intelligence, 2015, 37(9): 1904-1916.

[71] Gal R, Hochberg D C, Bermano A, et al. Swagan: A style-based wavelet-driven generative model[J]. ACM Transactions on Graphics (TOG), 2021, 40(4): 1-11.

[72] Sunkara R, Luo T. No more strided convolutions or pooling: A new CNN building block for low-resolution images and small objects[C]//Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Cham: Springer Nature Switzerland, 2022: 443-459.

[73] Tang C, Zhao Y, Wang G, et al. Sparse MLP for image recognition: Is self-attention really necessary?[C]//Proceedings of the AAAI conference on artificial intelligence. 2022, 36(2): 2344-2351.