Weighted Densenet Ensemble with Transformer Encoder and External Attention for Gastrointestinal Disease Classification

Gastrointestinal (GI) disease poses a higher health risk worldwide, necessitating timely and accurate diagnosis to prevent severe complications. Some traditional, manually performed analysis methods exist for the identification of these diseases, but they are extremely time-consuming and susceptible to mistakes. To address these challenges, deep learning- based approaches have garnered attention to automatically extract significant information and improve classification accuracy. This study proposed a weighted ensemble approach that utilizes a pretrained DenseNet201 model to classify five types of GI diseases using the publicly available Kvasir V2 dataset. The ensemble consists of two models: the first integrates a Transformer encoder with DenseNet201 for feature extraction, while the second combines DenseNet201 with an external attention mechanism. Both models use DenseNet201 as the feature map, which is subsequently passed through their respective attention layers. The final prediction is derived by computing a weighted sum of the last softmax predictions from both models. With 96.2% accuracy, 96.32% precision, and 96.2% recall, the suggested ensemble architecture surpasses the performance of individual models. This impressive performance highlights the potential of the proposed approach to significantly enhance classification accuracy, paving the way for improved diagnostics and earlier detection of gastrointestinal diseases. Early and precise diagnosis is crucial for effective patient management and treatment planning, ultimately leading to better clinical outcomes.