The presence of Microaneurysms (MAs) is the earliest detectable small abnormality of Diabetes Retinopathy (DR), a retinal disease that represents the leading cause of blindness among the middle-aged population globally. So the detection of MAs in fundus images is an important and challenging step for early diagnosis and prevention of critical health conditions. In particular, MAs are small vascular lesions consisting of swollen capillaries due to weakened vascular walls. Thus, retinal MAs can be associated with various ophthalmic and cardiovascular conditions. For instance, retinal MAs have been demonstrated as a risk factor for strokes. Therefore, it is crucial to detect the disease at its earliest stages to prevent its progression and consequent potential vision loss. However, MAs are a small target relative to the fundus image. Because the visual conditions are not ideal, MAs may present a low contrast with the background or may be affected by uneven illumination in the image. In addition, MAs may also be confused with other structures in the image, such as microbleeds, pigmentation changes, and dust particles in the fundus camera. Therefore, the automatic detection of MAs in fundus images is a significantly challenging task. The existing MA detection algorithm is difficult to achieve accurate detection of the lesion. Therefore, an improved faster-RCNN (Faster-RCNN-Pro) detection algorithm based on multi feature scale fusion is proposed. Firstly, the structure of feature extraction network and Region Proposal Network (RPN) are improved by using multi feature scale fusion to increase the utilization of micro target features; then, the quantization error introduced in the process of pooling the region of interest is eliminated by homogenizing and pooling the region of interest; finally, by redesigning the smooth L1 loss function in the loss function, a balanced L1 loss function is obtained to realize the optimization of the loss function, so as to effectively reduce the imbalance between large gradient difficult samples and small gradient easy samples, so that the model can be better trained. For the automatic detection of MAs in fundus images, the optimized Faster-RCNN network model is trained and tested on the Kaggle DR dataset, and compared with other methods. Based on the Kaggle DR dataset, the ablation experiment analysis verifies that the proposed improved Faster-RCNN-Pro based on multi-feature scale fusion can effectively improve the automatic detection performance of MAs. Specifically F-score is increased by 9.36% compared to that of Faster-RCNN. In addition, the performance of this method is compared to the automatic detection performance of MAs based on YOLO, CNN, and traditional methods including image processing and classifiers. The results demonstrates that the F-scores of the methods based on deep learning, including YOLO, CNN and the Faster-RCNN-Pro proposed in this study, are superior to the methods based on image processing and classifiers. F-score of traditional methods such as image processing and classifiers is low because traditional algorithms are easily limited by parameters. The lesion candidate regions of MAs extracted from a complex background, such as the fundus images, are more prone to interference and can not be excluded, and will eventually become FP, resulting in a low P value and affecting the F-score. Moreover, the two-stage model of Faster-RCNN presents a significantly better detection performance, such as F-score, due to the existence of RPN; the detection accuracy of the Faster-RCNN-Pro has been significantly improved. However, the deep neural network may also present the overfitting phenomenon, which may lead to some MAs not being detected. Therefore, the R values based on deep neural networks in previous studies are lower than those obtained by using traditional algorithms; however, Faster-RCNN-Pro can overcome this problem. So the proposed Faster-RCNN-Pro can accurately and effectively detect MAs in fundus images, demonstrating a better detection performance.