LungDetectNet: a multi-task deep learning framework with enhanced detection and descriptive capabilities

Artificial intelligence has significantly transformed medical image analysis, particularly in the early diagnosis of lung cancer from computed tomography (CT) scans. A key step in this diagnostic process is the accurate identification of lung nodules, which are primary indicators of potential malignancy, yet this identification task remains challenging due to their small size and subtle features. While efficient 3D object detection frameworks like MedYOLO offer a promising approach, their optimal architecture for this specific task is not well-established. To address this, we introduce LungDetectNet, a framework that advances the MedYOLO approach for 3D lung nodule detection. This advancement was achieved through a systematic investigation of the YOLO backbone’s evolution, which demonstrated a clear performance improvement corresponding with the integration of more advanced backbone architectural designs. In addition to detection, regression heads are integrated into the framework to predict seven descriptive attributes of each detected nodule: Subtlety, Sphericity, Margin, Lobulation, Spiculation, Texture, and Malignancy. On our newly established, challenging benchmark from the LIDC-IDRI dataset, our final model achieves a Mean Average Precision (MAP) of 0.793, with a precision of 0.896 and a recall of 0.664. For the multi-task regression objective, the model also shows strong performance, achieving an average Mean Absolute Error (MAE) of 0.516. These results demonstrate that LungDetectNet has the potential to enhance early-stage lung nodule detection and provide detailed diagnostic insights, therefore supporting clinical decision-making and improving patient care.