This study delves into the distinctive characteristics of lesions in Digital Retina (DR) images, aiming to construct and assess the performance of an intelligent diagnostic model tailored for scenarios involving limited sample sizes. The analysis encompasses a comprehensive examination of lesion types and recognition methodologies within DR images, with a specific emphasis on the noteworthy variations in shape, size, and color across diverse lesions. Subsequently, an intelligent diagnostic model grounded in the principles of small sample learning theory is meticulously developed. This model integrates advanced techniques, including meta-learning, transfer learning, and attention mechanisms, to augment its generalization capabilities and precision. The model’s training and validation phases are accompanied by meticulous data preprocessing, strategic training protocols, and rigorous validation methodologies. Experimental outcomes showcase the model’s exceptional proficiency in identifying various lesion types, with a notable highlight being its attainment of an AUC value of 0.94 in the discernment of microvascular abnormalities. To corroborate the model’s efficacy, rigorous t-tests are employed, and a comparative analysis with traditional DR image diagnostic approaches illuminates both the strengths and limitations of the proposed model.

DR images;lesion features;small sample intelligent diagnostic model

[1] Gao Z, Early Seizure Onset after Stroke: Analysis of Lesion Location Features in CT Images. China Contemporary Medicine, 2011, 18(05): 80-81. [2] Yang G, Huang J, Lesion Detection in Gastroscopic Images Based on Multi-Feature Fusion Matrix Decomposition. Modern Electronic Technology, 2022, 45(03): 69-72. [3] Hu W, Zou J, Guo Y, Zhang J, Wang B. Multidimensional Feature-based Method for Pathological Image Lesion Recognition. Computer Engineering and Applications, 2021, 57(06): 144-151. [4] Lin C, Ren C, Wang S. Analysis of Lesion Signal Characteristics and Distribution in MRI and DTI Images of Patients with Secondary Wallerian Degeneration due to Unilateral Infarction of the Pyramidal Tract. Chinese Journal of Endemic Diseases Prevention and Control, 2017, 32(03): 340. [5] Zhang J, Imaging Characteristics Analysis of Lesions in CT Images of Patients with Post-Stroke Dysphagia. Chinese Medical Frontier, 2011, 6(16): 68+36.

1.Supported by China University Innovation Fund《Research on Focal Feature Difference and Small Sample Intelligent Diagnosis Model of DR Images between Uyghur and Han Nationality in Southern Xinjiang》(2022IT100). 2.President's Fund of Tarim University《Research on Anatomical Structure of fundus Images and Segmentation of Lesions in Diabetic Retinopathy》 (TDZKSS202240).

DOI： https://doi.org/10.62639/sspjinss04.20240101

Published： 2024