Gastrointestinal diseases pose a significant threat to human health, with gastrointestinal tumors ranking among the most common and fatal conditions. Narrow Band Imaging (NBI), an advanced endoscopic method, plays a crucial role in diagnosing these diseases by producing detailed imaging data. However, the large volume of generated data often includes low-quality images, making it challenging to identify frames with diagnostic value. This underscores the importance of an efficient image quality control mechanism for data optimization. In this study, we introduce a non-reference image quality assessment (IQA) framework specifically designed for NBI endoscopy. The proposed two-stage approach employs deep learning for accurate evaluation. In the first stage, a patch-based classification model assesses image regions by leveraging local features extracted through a convolutional neural network. In the second stage, a breadth-first search algorithm combines these patch-level outcomes to generate an overall image quality score. The framework achieves outstanding performance, with precision and recall rates of 96% and 97%, respectively. Additionally, it reduces storage requirements by approximately 90% and adapts effectively to various application needs.