3D medical imaging is essential for the accurate diagnosis of complex pathologies, ranging from brain tumors to cardiovascular anomalies. However, manual analysis of these voluminous datasets by radiologists is time-consuming and prone to inter-observer variability. While Convolutional Neural Networks (CNNs) have demonstrated significant efficacy in 2D image analysis, their adaptation to 3D data presents both technical and clinical challenges, including the integration of expert knowledge and ensuring scalability in real-world environments. This paper addresses the classification of 3D medical images by proposing a methodology for designing responsible, auditable, and clinically relevant AI systems. The study explores various 3D CNN architectures both from transfer learning and from scratch for classifying pathologies in medical images, with a particular focus on lung CT scans. Key aspects include data preprocessing, model development, and performance evaluation using metrics such as accuracy, sensitivity, and specificity. The results indicate that carefully designed architectures and the application of transfer learning can achieve promising outcomes even with limited datasets, underscoring the importance of balancing model complexity with data availability and leveraging pre-existing knowledge.