眼底疾病是致盲的主要原因之一。借助光学相干层析成像技术（OCT），可实现早期眼底疾病的发现和及时治疗，是预防失明的有效手段。为缓解医生的阅片压力，计算机辅助诊断技术逐渐受到关注。然而，受限于眼底OCT数据的隐私性，计算机辅助技术的研究者无法获取数据来开展工作。针对该现状，检索梳理了8个免费的公开的眼底OCT数据库，对涉及的典型眼底疾病的OCT图像特征进行解释，并筛选出64篇基于这些数据做计算机辅助算法的文献，分类阐述了这些工作的贡献。为真正实现计算机辅助技术在眼底疾病早期诊断的临床应用，未来还可以从提高眼底OCT图像的高精度分类的稳定性可重复性和泛化能力、提高对眼底OCT图像的分割能力、提高计算机辅助算法的可解释性三方面进行努力。

基于胸部X光透射图像(DR)的肺部病灶识别与疾病诊断是临床的常规操作。对于肺结核患者而言, 其DR图像中的病灶区域与背景相融性高, 目标弥散严重且边缘形态极不规则, 严重干扰诊断的准确性。针对上述问题, 提出了一种融合肺炎影像学特征的肺结核病灶区域检测网络(TDT-Net), 利用肺结核和新冠肺炎同为呼吸道疾病且在DR图像上具有相似表征的特点, 借助大量肺炎DR数据, 通过迁移学习引入强相关特征以提高肺结核病灶的检测精度。TDT-Net结合Transformer和扩张残差技术, 提出上下文感知增强模块, 以强化迁移模型对全局信息的建模能力; 利用特征细化模块减少迁移过程中引入的冗余信息, 凸显强相关特征的表示。实验结果表明, 在TBX11K数据集上, 所提检测方法的平均准确度(AP)达到87.5%, 召回率(Recall)达到80.7%, 相较于YOLOV5和RetinaNet等网络有效提升了肺结核病灶的检测精度, 实现了更加准确的肺结核病灶定位和分类。

鲜红斑痣(port wine stains, PWS)是最常见的先天性皮肤微血管病变之一，PWS的病因是皮肤真皮层由浅至深的毛细血管畸形扩张。通常表现为面颈部粉色、红色和紫色斑片，随着年龄的增加，其逐渐加深和增厚，严重影响患者的生活质量。血管靶向光动力疗法 (vascular targeted photodynamic therapy, V-PDT) 可以选择性破坏病变血管，是目前国内治疗PWS的首选方法。V-PDT疗效与PWS病灶结构密切相关。PWS的病灶结构可通过活检或者无创光学诊断设备获取，主要包括表皮层黑色素含量、皮肤厚度及血管管径、深度和形态等。总结了目前常用的无创在体光学成像技术在PWS诊疗中的应用现状及PWS病灶结构特点对V-PDT疗效的影响，旨在为V-PDT精准及个性化治疗PWS提供参考。

针对小目标识别与分割问题，提出了一种基于双边融合的网络模型BFNet（Bilateral Fusion Network），该模型具有双分支结构。一支为具有较窄的通道和较浅的结构层，用于关注相邻像素点之间的联系。另一支引入RFB（Receptive Field Block）和DFB（Dense Fusion Block）两个模块，其具有较宽的通道和较深的结构层，可以获得高级语义的上下文信息。随后由一个引导聚合层融合两分支的特征表示。以息肉和皮肤病变区域为应用对象，使用三个公开的医学分割数据集来评估所提算法的性能。实验结果表明，在息肉和皮肤病变分割任务中，所提算法优于现有的医学图像分割算法。特别是在自动息肉检测Kvasir-SEG数据集中，所提算法的平均Dice和平均交并比分别达到了92.3%和86.2%，均比已有算法高。

High-grade squamous intraepithelial lesion (HSIL) is regarded as a serious precancerous state of cervix, and it is easy to progress into cervical invasive carcinoma which highlights the importance of earlier diagnosis and treatment of cervical lesions. Pathologists examine the biopsied cervical epithelial tissue through a microscope. The pathological examination will take a long time and sometimes results in high inter- and intra-observer variability in outcomes. Polarization imaging techniques have broad application prospects for biomedical diagnosis such as breast, liver, colon, thyroid and so on. In our team, we have derived polarimetry feature parameters (PFPs) to characterize microstructural features in histological sections of breast tissues, and the accuracy for PFPs ranges from 0.82 to 0.91. Therefore, the aim of this paper is to distinguish automatically microstructural features between HSIL and cervical squamous cell carcinoma (CSCC) by means of polarization imaging techniques, and try to provide quantitative reference index for pathological diagnosis which can alleviate the workload of pathologists. Polarization images of the H&E stained histological slices were obtained by Mueller matrix microscope. The typical pathological structure area was labeled by two experienced pathologists. Calculate the polarimetry basis parameter (PBP) statistics for this region. The PBP statistics (stat PBPs) are screened by mutual information (MI) method. The training method is based on a linear discriminant analysis (LDA) classifier which finds the most simplified linear combination from these stat PBPs and the accuracy remains constant to characterize the specific microstructural feature quantitatively in cervical squamous epithelium. We present results from 37 clinical patients with analysis regions of cervical squamous epithelium. The accuracy of PFP for recognizing HSIL and CSCC was 83.8% and 87.5%, respectively. This work demonstrates the ability of PFP to quantitatively characterize the cervical squamous epithelial lesions in the H&E pathological sections. Significance: Polarization detection technology provides an e±cient method for digital pathological diagnosis and points out a new way for automatic screening of pathological sections.

Recently, research has been conducted to assist in the processing and analysis of histopathological images using machine learning algorithms. In this study, we established machine learning-based algorithms to detect photothrombotic lesions in histological images of photothrombosis-induced rabbit brains. Six machine learning-based algorithms for binary classification were applied, and the accuracies were compared to classify normal tissues and photothrombotic lesions. The lesion classification model consisting of a 3-layered neural network with a rectified linear unit (ReLU) activation function, Xavier initialization, and Adam optimization using datasets with a unit size of 128 × 128 pixels yielded the highest accuracy (0.975). In the validation using the tested histological images, it was confirmed that the model could identify regions where brain damage occurred due to photochemical ischemic stroke. Through the development of machine learning-based photothrombotic lesion classi- fication models and performance comparisons, we confirmed that machine learning algorithms have the potential to be utilized in histopathology and various medical diagnostic techniques.

针对黑色素瘤与非黑色素瘤在视觉上相似度高、颜色多样、边缘模糊、异物遮挡等情况而导致皮肤病变分割效果差的问题,提出一种基于U型结构的上下文编码解码网络,通过采用高效双通道注意力机制模块和空洞空间金字塔池化模块来捕获更多的语义信息与空间信息,以提高皮肤病变的分割精度。在ISIC 2017皮肤镜图像数据集上进行训练和测试,实验结果表明,本文算法分割结果的相似度系数(Dice_Coefficient)高达88.74%,比目前主流语义分割网络模型DeepLab V3 Plus高3.15个百分点,比医学领域经典U-Net网络高9.93个百分点,且运行速度快和稳定性好,能对黑色素瘤实施有效分割,分割图像边缘连续、轮廓清晰,在定量分析和识别中使用效果良好。

Computer-assisted cervical screening is an effective method to save the doctors' workload and improve their work e±ciency. Usually, the correct classification of cervical cells depends on the nuclear segmentation effect and the extraction of nuclear features. However, the precise nucleus segmentation remains a huge challenge, especially for densely distributed nucleus. Moreover, previous cellular classification methods are mostly based on morphological features of nucleus size or color. Those individual features can make accurate classification for severe lesions, but not for mild lesions. In this paper, we propose an accurate instance segmentation algorithm and propose cognition-based features to identify cervical cancer cells. Different from previous individual nucleus features, we also propose population features and cognition-based features according to the Bethesda System (TBS) for reporting cervical cytology and the diagnostic experience of the cytologists. The results showed that the segmentation achieves better success in complex situations than that by traditional segmentation algorithms. Besides, the cell classification via cognition-based features also help us find out more about less severe lesions' nuclei than that based on conventional features of individual nucleus, meaning an improvement of classification accuracy for cervical screening.

农作物生长发育过程中经常会遭到病虫害等外界因素侵染, 如果不能实施有效的监测诊断和科学的防治, 极易引起农药喷洒不当或过量, 不仅会影响作物的产量和种植户的经济效益, 还会造成严重的环境污染。 近年在广西大棚厚皮甜瓜上发生了一种严重的由瓜类尾孢(Cercospora citrullina)引起的甜瓜叶斑病, 导致甜瓜减产和种植户的经济损失。 故此应用高光谱成像开展甜瓜叶片的尾孢叶斑病检测, 获取健康甜瓜叶片和受瓜类尾孢感染的具有不同病变程度的甜瓜叶片在380~1 000和900~1 700 nm的高光谱图像, 选取感兴趣区域并获取相应的平均光谱反射率, 比较发现健康叶片和不同病变程度叶片染病区域的平均反射率差异显著。 在540 nm处附近, 健康叶片和病变程度轻微的叶片的光谱具备波峰形态, 随着病变程度增加, 波峰逐渐消失; 在700~750 nm处附近, 叶片反射率曲线急剧上升, 出现绿色植物光谱曲线显著的“红边效应”特征; 750~900 nm范围, 健康叶片与轻微病变区域的光谱反射率变化趋于平稳, 而其他病变区域的反射率呈上升趋势, 且健康叶片的反射率高于病变区域, 反射率随病变程度增加而下降, 这一变化规律一直持续到近红外波段的900~1 350 nm范围。 运用主成分分析、 最小噪声分离法观察叶片早期病变的特征, 经主成分分析和最小噪声分离法处理后, 特别是对于早期病变, 样本受感染后发病的区域更为明显。 基于高光谱图像提取的前三个主成分得分绘制三维散点图, 虽然不同病变程度的部分样本有重叠, 但病变样本与健康样本的分布区分明显。 应用K-近邻法和支持向量机方法建立叶片病变判别模型, 结果显示: KNN模型对健康样本测试集判别率为98.7%, 病变样本的判别率随病变程度加重而逐渐升高; 对病变程度较轻样本, 支持向量机模型相比于KNN模型而言, 判别正确率更高、 分类效果更好; 总体上, 高光谱图像对健康样本的判别率较高(>97%), 可用于健康样本与病变样本的识别, 但对不同病变程度的区分效果欠佳。 研究结果表明, 高光谱成像可用于甜瓜尾孢叶斑病的检测, 对不同病变程度的区分效果仍有待提高。