An accurate assessment of the morphological characteristics of the Optic Disc (OD) is essential for the diagnosis of various retinal disorders. It is necessary to segment the OD precisely to detect structural OD changes associated with visual field loss. Although deep learning models are effective for this task, they require extensive labeled datasets, which can be time-consuming and costly. Furthermore, fundus images have multi-scale features, making segmentation challenging. In this study, we present a semi-supervised and transfer learning approach for OD segmentation. Our approach utilizes an improved Dilated Convolutional AutoEncoder (DCAE) and a pre-trained modified U-Net to segment the OD. The DCAE segments the OD using feature similarity from unlabeled images in the Messidor dataset and saves the learned weights. Transfer learning is then applied to reuse the model weights in the U-Net, accelerating training on small datasets such as Drions-DB and Drishti-GS. The network architecture was modified by increasing the layers from 8 to 128 and halving the feature map length and width. To address the multi-scale challenge without inflating the model parameters, we introduce the Dilated Hierarchical Feature Extraction Module (DHFEM), a convolutional module capable of achieving multi-scale feature extraction without increasing model parameters. Additionally, DHFEM incorporates convolutional layers with varying receptive fields, further enhancing the network ability to extract features across multiple scales. Our OD segmentation method outperforms existing algorithms with reduced parameter quantities of 0.4 M. The mean Intersection over Union (mIoU) values are 0.9383 and 0.9629 and inference times of 45 ms and 40 ms for the Drions-DB and Drishti-GS datasets, respectively.