Développement d’un modèle d’intelligence artificielle pour la détection précoce des nodules pulmonaires cancéreux

Abstract

Lung cancer is the leading cause of cancer-related mortality worldwide and a major public health concern in Algeria, typically characterized by a silent progression that leads to late-stage diagnosis. This research work presents the design, implementation, and rigorous evaluation of an innovative artificial intelligence system dedicated to the automated and early detection of pulmonary nodules from chest computed tomography (CT) scans, combining technological expertise with clinical analysis. The proposed methodology is based on a semi-3D (2.5D) hybrid architecture executed within the Kaggle Cloud environment using Python 3.12 (TensorFlow, Keras, SimpleITK). The network couples local feature extraction via a MobileNetV2 backbone (leveraging partial fine-tuning on the last 20 layers) with global contextual modeling through a 4-head Multi-Head Attention mechanism. The model was trained on a merged and standardized dataset consisting of 3,645 isotropic patches (1 mm³) from the benchmark LUNA16 and RadGenome-CT datasets, cleaned using standard deviation analysis and normalized (clipping from -1000 to 400 HU). To mitigate the substantial class imbalance (2,729 healthy vs. 916 nodules), a strict cost-sensitive Class Weighting strategy and an on-the-fly data augmentation pipeline were successfully deployed. Experimental evaluation on a 20% stratified validation set demonstrates the system's robustness and high discriminative capabilities. The model achieved an outstanding Area Under the ROC Curve (AUC) of 0.96 and an Average Precision (AP) of 0.90 on the Precision-Recall curve. Detailed classification metrics reveal excellent clinical sensitivity, yielding a recall of 87% for the nodule class (160 true positives and only 23 false negatives), thereby drastically reducing the risk of diagnostic omissions. Specificity is highlighted by a 91% recall and a 96% precision for the healthy class (496 true negatives). FROC curve analysis confirms a sharp early ascent, reaching a 0.90 sensitivity with a very low number of false positives per scan. Furthermore, the black-box model’s clinical explainability was ensured via occlusion sensitivity activation maps (heatmaps), visually validating that the network focuses strictly on the intended pathological structures. This system stands out as a highly reliable and powerful computer-aided pre-diagnosis tool capable of supporting radiologists in clinical workflows.