In the era of artificial intelligence (AI), particularly with the progress in convolutional neural networks (CNNs) and the digital transformation of endoscopic images, computer-aided diagnosis for gastrointestinal diseases has become feasible. However, many existing models exhibit significant computational complexity and involve a high number of parameters, rendering them less feasible for deployment on edge devices. The current emphasis on edge computing underscores the need for cost-effective solutions in the realm of gastrointestinal diagnosis. In addressing these challenges, our paper introduces “GastricNet,” a CNN model designed to be efficient and lightweight for the detection of gastrointestinal diseases. GastricNet incorporates several novel components. Firstly, it utilizes a lightweight residual block (Lite-ResB) to optimize computing costs and mitigate issues related to vanishing and exploding gradients. Secondly, the ConvMixer includes both Depthwise Convolution and Pointwise Convolution, separating spatial and channel dimension mixing to enhance efficiency and reduce redundancy in features. Thirdly, the hypercolumn technique is implemented to preserve local discriminative features from different levels of GastricNet. Lastly, the Squeeze and Excitation (SE) block is employed to enhance feature representation, dynamically adjusting the importance of different channels. Moreover, we utilize knowledge distillation (KD) to enhance the model’s representation capabilities, thereby improving performance while still maintaining the computational efficiency. GastricNet, along with 11 state-of-the-art (SOTA) pre-trained models, has been successfully deployed on two mobile devices, showcasing its ability to operate efficiently on lightweight and portable processors. Compared to state-of-the-art pre-trained models, GastricNet demonstrates exceptional performance with an accuracy of 98.37%, a compact model size of just 1.1 MB, only 66.2k parameters, and lower computational demands, leading to faster image classification. This work also conducts a comprehensive ablation study, elucidating the impact of various components on GastricNet’s performance. Experimental results on an endoscopic image dataset demonstrate the superior performance of GastricNet on edge devices, exhibiting higher accuracy, lower complexity, and reduced memory requirements. This underscores its efficiency concerning both accuracy and a lightweight design, positioning it as a promising choice for deployment on mobile edge devices.
Asif et al. (Fri,) studied this question.