HUCSR-Net: Automated Pulmonary Embolism Detection Using 3D Deep Learning

Background Pulmonary embolism (PE) remains a leading cause of cardiovascular mortality and a major diagnostic challenge in emergency settings worldwide. Although computed tomography pulmonary angiography (CTPA) is the reference standard, its interpretation is time-consuming, subject to inter-observer variability, and dependent on subspecialist expertise, potentially delaying life-saving treatment. Methods We developed and validated HUCSR-Net, a fully automated 3D convolutional neural network for PE detection using 128,484 imaging studies from 86 patients at Hospital Universitario Clínica San Rafael (Bogotá, Colombia). A modern R(2 + 1)D-18 architecture, pre-trained on Kinetics-400, was adapted to single-channel input by averaging the RGB weights of the first convolutional layer. The classification head was replaced by a dropout layer (p = 0.6) followed by a single linear unit. The model was trained from scratch with patient-level 5-fold cross-validation, using AdamW optimiser and binary cross-entropy with logits loss. All experiments were conducted on a workstation equipped with an NVIDIA GeForce RTX 5070 Ti GPU (16 GB VRAM) and an AMD Ryzen 9 9950X processor. Performance was evaluated by the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and F1-score. Results After several experiments, the optimal model was obtained with a validation AUC of 0.7026, sensitivity of 0.8592, specificity of 0.2673, precision of 0.4519, and F1 score of 0.592. The Matthews correlation coefficient was 0.1515, and the area under the precision-recall curve was 0.5907, confirming solid discriminatory performance despite the limited size of the cohort and validation losses of 0.7989, respectively. Conclusions A deep 3D convolutional network trained from scratch on a modest single-centre cohort can achieve diagnostic performance comparable to published multi-thousand-patient studies relying on large public datasets. These results demonstrate the feasibility of clinically useful automated PE detection in resource-constrained settings and support the integration of such systems as decision-support tools for radiologists.