Implementation of an advanced lung nodule classification system using optimized ConvMixer and AdamW-based CNN architecture

Early detection of lung cancer has the potential to save lives, provided that it exhibits reliable, smooth, and efficient performance. Many recent cases have been diagnosed with lung cancer; however, inadequate and delayed detection methods have increased mortality rates. Advanced automated systems have been developed to analyze lung CT images and identify critical abnormalities within the lungs. Nevertheless, the complexity of lesions and nodules' morphology presents a formidable challenge, hindering the attainment of desired outcomes. Convolutional neural networks (CNNs) have emerged as a promising approach in recent studies, demonstrating exceptional clarity and confidence in identifying various lung structures within CT scans. Nevertheless, traditional CNN architectures possess certain limitations, such as a pyramidal distribution of neural network layers that extracts an increasing number of features while reducing resolution. Consequently, these operations sometimes fail to deliver the desired level of image analysis performance due to characteristics like local receptive fields, data inefficiency, and unstable behaviour, resulting in suboptimal classification results. To address these limitations, this research paper introduces the ConvMixer-CNN (CMCNN) model, which leverages an isotropic vision architecture, a global attention mechanism, and different spatial and channel dimensions mixing techniques. The model also incorporates the Gaussian Error Linear Unit (GELU) as an activation function, enabling smoother operations. Furthermore, the proposed technique replaces the commonly used Adam optimizer with the AdamW (Adaptive Movement Estimation with Weight Decay) optimizer. Comparative evaluations against other recent promising CNN-based techniques for lung nodule classification demonstrate that the proposed approach performs favourably. When evaluated using various metrics, including accuracy (97.02%), recall (93.12%), precision (89.57%), and F1-score (87.36%), the proposed technique exhibits superior results. To further enhance the technique, future research could encompass experiments using multiple datasets and exploring advanced optimizer functions for training, aiming to generate a more standardized model for the lung nodule classification task.