Abstract
<title>Abstract</title> <p>Precise recognition and classification of pathological features in medical imaging are fundamental to the development of automated diagnostic systems and computer-aided clinical decision-making. While Convolutional Neural Networks (CNNs) utilizing standard activation functions such as ReLU, Leaky ReLU, and ELU have demonstrated significant success, a critical research gap remains in exploring alternative functions that can better capture the complex, non-linear patterns inherent in medical imagery. In this work, we address this gap by proposing a novel sinusoidal-based activation function named the Rectified Sine Unit (RSU). The proposed RSU is evaluated across four diverse medical datasets: Brain Cancer (MRI), Lung Cancer (CT), Colon Cancer (Histopathology), and Breast Cancer. Experimental findings validated that CNN models integrated with the proposed activation function consistently outperform or remain highly competitive with the current state of the art. Specifically, RSU achieved a peak accuracy of 85.31% in Brain Cancer classification (outperforming the ReLU-based model at 83.00%), a near-perfect 99.55% in Lung Cancer (surpassing Leaky ReLU’s 98.64%), and 97.88% in Breast Cancer (exceeding Leaky ReLU’s 97.27%). In the Colon Cancer dataset, RSU demonstrated high robustness with a test accuracy of 99.77%. These results demonstrate that the periodic nature of the proposed activation function facilitates the exploitation of more intricate spatial features, resulting in an enhanced understanding and feature extraction at deeper network levels. This study contributes a compelling case for the adoption of new activation functions to advance the search for efficient and reliable CNNs in the field of automated medical diagnosis.</p>