In this paper a neural network method for iris image key points detection based on handcrafted features using the Key.Net architecture is proposed. Due to the use of the handcrafted features in CNN, the proposed method combines the robustness of classical key points detection methods and high accuracy of neural networks. Additional Hermite-based convolutional filters are integrated into the network to improve keypoint localization. A synthetic dataset is generated from normalized iris images using geometric and photometric transformations. Matching of iris image key points is performed using HardNet descriptors, followed by geometric filtering and confidence-based ranking. Experimental evaluation demonstrates the robustness of the proposed method to the presence of eyelids and eyelashes without using any segmentation masks. The proposed approach achieves an Equal Error Rate (EER) value of 0.096% on the CASIA-IrisV4-Interval database. These results show the potential for the combination of handcrafted filtering with deep learning for accurate and interpretable iris recognition.