Super-Efficient Lung Nodule Classification Using Massive-Training Artificial Neural Network (MTANN) Compact Model on LIDC-IDRI Database

Accurately diagnosing lung cancer relies on analyzing hundreds of images in a single examination, placing a significant workload on radiologists. In radiation treatment planning, tailoring the prescribed dose distribution to each tumor area is critical. However, manual tumor region demarcation has proven laborious and time-consuming, with inconsistencies potentially leading to intra- and inter-session variations in tumor region definition. Amidst the vast landscape of deep learning, massive-training artificial neural networks (MTANN) stand out for their accuracy in image pattern recognition. In this paper, 3D MTANN is used to classify nodules based on their level of malignancy on lung CT images in the LIDC-IDRI database. Detection is fully automated and evaluated using the AUC of the ROC curve. By adjusting the 3D MTANN parameters, our method achieved a maximum AUC of 0.91 with efficient inference capabilities. Experimental results demonstrate the reliability of MTANN in differentiating benign and malignant nodules from thoracic CT images with better accuracy than state-of-the-art methods. The proposed method successfully achieved its goal of determining lung cancer by maintaining a high level of differentiation while reducing the proportion of false positives. Our 3D MTANN-based AI system with efficient inference capabilities would be useful in assisting radiologists in their diagnosis of lung nodules effectively.