Introduction: Lung cancer is one of the leading causes of cancer mortality worldwide (MS, 2024). The area of digital image processing and analysis is one of the most important fields in medical science due to the rapid and continuous progress in medical image visualization and advances in computer-aided diagnostic methods and image-guided therapies. This area has been essential for early detection, diagnosis, and assessment of cancer treatment response. The challenge is to effectively process and analyze medical images to extract, quantify, and interpret this information to gain understanding and insight into the structure and function of organs and systems (Silva et al., 2021). Objective: The objective of this work is to implement and critically evaluate a pulmonary nodule classification pipeline based on computed tomography images using deep learning techniques, specifically deep neural networks. Material and Methods: The proposed pipeline consists of two sequential stages: a semantic segmentation model using the nnU-Net framework to identify and delineate regions of interest (ROIs) as nodule candidates, and secondly, a machine learning model based on the volumetric Inception architecture. The system was trained and validated using two datasets of low-dose tomographic images annotated by specialists: a combined dataset from the public LIDC-IDRI repositories with exams from 1,018 patients and the Duke University repository with 2,061. The data are publicly accessible with patient data anonymization. The implementation sequence consisted of image preprocessing for selecting areas of interest corresponding to the lung parenchyma and mediastinum with 2.5D volume; semantic segmentation using the U-Net architecture under the nnU-Net v2 framework for generating nodule candidates suggestive of lung cancer; and finally, the binary classification of the latter, using an Inception convolutional network. Results: The segmentation stage achieved significant performance, with a mean Dice Similarity Coefficient (Dice) of 0.846 on the test set, which was made possible by applying data mass reduction techniques from 3D to extended 2.5D. The isolated classifier demonstrated high discriminative capacity, with an area under the ROC curve (AUC) of 0.933. The evaluation of the integrated pipeline revealed a sensitivity of 0.89 and a specificity of 0.97. In-depth analysis of the results identifies the quality of segmentation as the main performance bottleneck, demonstrating that the sensitivity of the end-to-end system is intrinsically limited by the capacity of the initial candidate generation stage. Conclusions: The two-stage hybrid architecture proved to be a methodologically solid strategy for balancing sensitivity and specificity. The system demonstrated robust false-positive filtration (averaging 0.15 false positives per image), validating its potential to mitigate radiological "alarm fatigue." However, the segmentation stage acts as a critical "performance ceiling," where missed lesions become unrecoverable, intrinsically limiting the global sensitivity despite the high precision of the subsequent classifier. Furthermore, clinical viability is currently hindered by significant infrastructure barriers, specifically the need for high-performance hardware, and the scarcity of curated local datasets required to generalize models to the national epidemiological profile. Future efforts must prioritize overcoming the segmentation sensitivity bottleneck and investing in native data curation to ensure applicability in clinical practice.