Due to its complexity, the Arabic language and its extensions build a fertile field of research in the field of artificial intelligence in general and optical character recognition (OCR) specifically. There are several languages that use the Arabic alphabet in their manuscripts. These languages innovated new letters to pronounce sounds not found in the Arabic language. These letters are called 'Arabic-derived letters'. To enrich the Arabic language, we can use these letters to know the true pronunciation of intrusive words in the Arabic language. This article deals with the Arabic-derived letters (ADL) dataset. It is a new dataset that consists of 55440 scanned images of papers written by 30 participants of different ages, with a data augmentation technique to increase the number of images. This study aims to evaluate and compare the effectiveness of different convolutional neural network architectures for ADL recognition, focusing on accuracy, robustness, and generalization capability. Three architectures were implemented: LeNet, a simplified ResNet model with residual blocks, and a deep VGG-Like network. Training was limited to 40 epochs with early stopping after 5 epochs without improvement. Experimental results show that the VGG-Like model achieves the best performance with 99.61% accuracy in validation, closely followed by ResNet with an accuracy of 98.98%. In contrast, LeNet performs less efficiently by 96.43%. These results clearly demonstrate that modern and deep architectures provide better accuracy and robustness for the classification of handwritten characters.

Skin cancer is the most fatal type of cancer worldwide. Prognosis is generally much better with early detection and effective treatment relies on an accurate diagnosis of skin lesions. Despite progress in deep learning, there is still a challenge to provide accurate segmentation of dermoscopic skin lesions due to image variations stemming from different illuminative input parameters, different resolutions/image sizes, image artifacts and varying skins. These differences hinder state-of-the-art models from learning pixel-precise lesion boundaries or more contextual features, consequently impairing their generalization and clinical applicability. We introduce a new architecture called Dual Focus Attention Block UNet (DFAB-UNet) with intrinsic capability of local and global feature extraction tasks. We trained and tested the proposed method on two commonly used sets of dermoscopy images, PH2 and HAM10000 with respect to six different parameters, namely, Accuracy (Acc), Precision (Pre), Sensitivity (Sen), Dice Similarity Coefficient (DSC), Intersection over Union (IoU), and specificity (Spe) for comparative analysis Results. DFAB-UNET achieved an Acc of 97.24%, a DSC of 94.30%, an IoU of 89.21% and a Spe of 98.95% on the PH2 dataset. On the HAM10000, we achieved an Acc of 96.20%, DSC of 92.79%, IoU of 86.55% and Spe of 97.60%. These results show that the model generalizes across datasets by introducing boundary precision represented in the loss together with contextual lesion information. The DFAB-UNet model achieves high performance on segmentation, indicating the potential use and application of dermal disease research in clinic practice.