Performance of the hdaru-net based deep learning approach on medical images: A study on adenoid hypertrophy and skin lesions

The accurate segmentation of medical conditions such as skin lesions and adenoid hypertrophy, as well as the classification of adenoid hypertrophy, plays a critical role in patient care. However, traditional diagnostic methods such as manual image analysis and visual examination pose significant challenges that limit diagnostic accuracy. Since these conventional techniques are typically performed by experts through endoscopy or radiographic evaluation, they are not only time-consuming but also prone to human error. This leads to diagnostic delays, repeated examinations, and increased costs due to the necessity of expert involvement. In recent years, advancements in artificial intelligence (AI) and deep learning have introduced automated approaches that can enhance diagnostic efficiency, particularly in medical image analysis. However, traditional deep learning models still struggle to handle complex variations such as noise in images, variations in lesion appearance, and irregular contrast. This study aims to overcome these limitations by implementing deep learning approaches, including the Hybrid Dense Attention Residual (HDARU-Net) model for segmentation and ensemble deep learning-based classification models for classification. Using a unique adenoid dataset obtained from Batman Training and Research Hospital and the International Skin Image Collaboration (ISIC) skin lesion dataset, the study comparatively examines the performance of the HDARU-Net model against deep learning-based segmentation models such as U-Net, U-Net++, and SegNet. The proposed HDARU-Net model enables more precise delineation of relevant regions in medical images, which is crucial for ensuring more reliable diagnoses and consistency in patient cases. The results demonstrate that integrating segmentation with classification tasks significantly improves diagnostic accuracy, particularly for complex skin lesions and adenoid hypertrophy cases. This study reveals that the HDARU-Net model not only reduces diagnostic time but also minimizes repetitive procedures requiring manual analysis, thereby lowering economic costs. The findings have significant implications for accelerating the diagnostic process in clinical settings, supporting timely medical interventions, and improving patient outcomes. This thesis study contributes to AI-based medical diagnostics by providing a robust framework that can also be applied to the diagnosis of other complex medical conditions.