Deep learning (DL) technology has been the best way to execute class over the most recent couple of years. These techniques were extended more specifically to retinal blood vessel segmentation, classifications, and predictions. One of the most significant technologies has been a deep learning technique, U-Net. During this research, we suggested improving the segmentation of retinal images in U-Net, bi-directional ConvLSTM U-Net (BiDCU-Net) with fully connected convolutional layers, like absolute U-Net, bi-directionally Convolutional LSTM (BiConvLSTM) preferences as well as the fully connected layers method. Rather than a basic link in the skip connection of U-Net, we utilize BiConvLSTM to join the feature maps extricated from the comparing encoding way and the past decoding up-convolutional layer in a straightforward manner. For enhancing the distribution and empowerment of highlighting, we use fully connected convolutional layers during the last encoding layer. Consequently, by using batch normalization (BN), we can speed up the configuration of this proposed network. Three recognized datasets were evaluated on the proposed technique: DRIVE, STARE and CHASED DB1 data sets. The visual as well as the quantitative findings show the strength of the approach suggested. This proposed methodology was carried out with the appropriate detailed measurements, accuracy, F1 score, sensitivity, and specificity in DRIVE, 0.9732, 0.8385, 0.8256, and 0.9868 in CHASE, 0.9744, 0.8194, 0.8392 and 0.9845 in STARE, 0.9733, 0.823, 0.8212 and 0.9857, respectively. Furthermore, we state that the approach is better than three similar techniques.