A Physiologically Informed Three-Dimensional Convolutional Neural Network to Aid Diagnosis of Non-Small Cell Lung Cancer

Lung cancer can be divided into two main categories, i.e., either small cell lung cancer (SCLC) or Non-Small Cell Lung Cancer (NSCLC), which accounts for 85% of lung cancers. NSCLC is a disease wherein malignant cancer cells form in the lung tissue. The five-year survival rate decreases as the NSCLC cancer becomes more advanced, from 40% for stage I to only 1% for stage IV; thus, a vital challenge to overcome is the early detection of NSCLC to achieve a higher chance of survival. As sonographers and radiologists are faced with the problem of being overworked with many time-consuming duties, it is important to consider new systems that could help in improving their effectiveness and efficiency while maintaining a clinically acceptable level of accuracy and reliability. Artificial Intelligence (AI) paves the way to allow computerized algorithms to perform tasks that would normally require only expert human involvement. In this report, a new AI-driven system is proposed to aid radiologists in the diagnosis of NSCLC, thus improving its specificity by design, while minimizing false positives as much as possible. It is worth noting that many clinical sub-fields have already started using AI to aid decision making, such as in breast cancer screening. Computed Tomography (CT) scans contain crucial volumetric, morphological lung tissue-related information that is only partly used to inform a diagnosis of NSCLC. In CT scans, each voxel can be converted into its corresponding Hounsfield Unit (HU) values, which are used to estimate tissue density; via this complementary information, areas of abnormal density or high attenuation can be identified, thus leading an AI-driven model to achieve a higher reliability in aiding diagnosis of NSCLC. Therefore, a three-dimensional AI-driven system may lead to an earlier detection of NSCLC, by complementing radiologists' expertise, experience, and decision making. This study leverages a thresholding-based segmentation algorithm to identify the region of interest defining the lungs by analysing each voxel's HU. Subsequently, an AI-driven model is proposed to analyse all CT scans of a patient to identify NSCLC earlier. This model is based on a three-dimensional convolutional neural network, trained on volumetric information of 90 images and validated on 40 images. The CT scan dataset has metadata for all patients and allows to compare the classification models predictions to the known or ground-truth diagnostic labels of the patients. The model led to 100
^%
classification accuracy and an area under the precision-recall curve of 1 (100%) over 130 subjects (70 patients with NSCLC, 60 healthy controls). Further work is underway to train, validate, and test the model on a larger dataset and assess its performance more comprehensively, and such minimum sample sizes will be determined statistically based on a power analysis, so that the results will be confidently comparable against the clinical diagnostic gold-standard labels.