Artificial Intelligence in Lung Cancer Diagnosis: “SYNERGY-NET” in Campania FESR-POR (European Fund of Regional Development—Regional Operative Program) Research Project

The screening protocols organized in recent decades have made it possible to diagnose an ever-increasing number of lung cancers at an early stage in order to allow their surgical removal, the only strategy with therapeutic purposes. However, only 60% of nodules identified on Low Dose Computed Tomography (LDCT) of the chest are neoplastic. Despite the improvement of imaging techniques and the use of PET, the differential diagnosis between malignant and benign lesions still remains a challenge, especially in the case of peripheral and subcentimetric lesions. The objective of the project was to evaluate whether artificial intelligence can provide additional information compared to standard techniques in the differential diagnosis of malignant neoplastic nodules. The following clinical data of lung cancer patients were taken into consideration: age, sex, comorbidity, tumor characteristics (histology, size, stage), type of operation, post-operative complications, and clinical follow-up. Subsequently, radiological images of patients with lung injury were analyzed. The Region of Interest of the lesion (ROI) was analyzed in order to provide artificial intelligence (CNNs) with information for the creation of an algorithm that can facilitate the differential diagnosis between malignant and benign lung lesions. The study of lung images (acquired via computed tomography) on a sample of 330 LDCTs of the lung, in order to carry out an “intelligent” detection (by means of bounding boxes) of the suspicious tissues produced the following results: 230 True Positive (TP), 103 False Negative (FN), with a Sensitivity (SEN) of 61%. The study of lung images (acquired via computed tomography) on a sample of 330 LDCTs of the lung, in order to carry out an “intelligent” detection (using bounding boxes) of the suspicious tissues produced the following results: 230 True Positive (TP), 103 False Negatives (FN), with a Sensitivity (SEN) of 61% and with 63% harmonic mean of precision and recall (F1). Thanks to the transfer learning methodology, the next step was to provide a system capable of carrying out the classification of the previously identified ROI. In particular, the MobileNet CNN is used to perform the diagnosis of the selected region. To make up for the small amount of datasets the network involved is pre-trained on the ImageNet dataset. The process of fine-tuning involves parameters to align with particular objectives and addressing the capabilities for problem resolution. The positive class is assigned to lesions declared malignant, while the negative label is used for benign ones. The performance of the system created in this way was significantly improved and evaluated in terms of overall Accuracy 89.47%, Specificity 76.92%, Sensitivity 93.19%, aligning with those of the literature and offering interesting insights into an algorithm capable of adapt native images in different formats to the created CAD-CNNs.