Deep learning, particularly Convolutional Neural Networks (CNNs), has transformed artificial intelligence (AI) applications in medical image analysis, enabling significant advancements in diagnosis and treatment. CNNs excel at extracting complex patterns from medical images, making them valuable tools for identifying diseases like tuberculosis (TB). However, despite their benefits, CNN models are both computationally intensive and resource-demanding. Training deep learning networks typically requires substantial memory, processing power, and energy consumption, making them impractical for deployment on devices with limited computational resources. The field continues to progress, yet there is still a notable gap in research dedicated to creating efficient and practical solutions for real-world healthcare applications, particularly in low-resource settings. This study addresses these challenges by developing a lightweight CNN for TB detection, incorporating pruning and quantization to reduce training time and model size, allowing rapid inference without loss of accuracy. The model is suitable for mobile and portable devices, especially in areas with limited access to advanced diagnostics and specialists. To evaluate performance, pruning and quantization are applied to established CNN architectures, assessed on accuracy and model size to analyze efficiency-performance trade-offs comprehensively. Notably, our model achieved an overall accuracy of ~98%, with a compact model size of 7.3 MB. This constitutes approximately ten times (10×) reduction in size compared to the baseline models while maintaining a comparable level of diagnostic accuracy. These optimizations make the model suitable for mobile and low-power devices, potentially transforming TB diagnostics in remote or underserved areas.