Automatically predicting lung tumor invasiveness using deep neural networks

Early lung cancer invasive detection is important for further treatment and saving lives. In clinical practice, lung tumor invasiveness (LTI) detection is very challenging, imaging-based automatic prediction algorithms offer a non-invasive approach. However, the lack of publicly available datasets and the imbalance of clinical categories are key issues limiting the development of automatic predictive methods. To address the above issues, a large well-labeled high-quality computed tomography dataset was collected from 804 patients, and each sample was labeled with a binary classification label according to the gold standard pathological report after surgery. Then, a novel artificial system, lung tumor invasiveness prediction neural network (LTI-Net), was proposed to perform the binary classification of lung tumors by solving the class imbalance problem and improving the performance diagnosis under such imbalance settings. We adopted a three-dimensional residual neural network as the backbone architecture to effectively captures intra-tumor heterogeneity through scanning the distribution changes of CT values in lesion regions in imaging data. Additionally, we introduced a novel surrogate function to approximate the area under the curve (AUC) metric. By leveraging both positive and negative sample pairs during the training process, this formulation enhances discriminative feature extraction while maintaining stable optimization dynamics. Comprehensive experiments on our collected dataset demonstrated the potential of our LTI-Net method. LTI-Net improved the score of the harmonic mean of true positives rate and true negatives rate (HMoPN) significantly when compared to the current state-of-the-art methods and improved 2.92% of the HMoPN score in different imbalanced settings.