Application of Artificial Intelligence for Efficient Tumor Volume Segmentation of Lung Cancers in the 18FDG PET-CT Modality

Broncho-pulmonary cancer (BPC) remains the primary cause of morbidity from malignant tumors in Tunisia. By being the first killer of all cancers, bronchial cancer bears witness to a disastrous and bleak prognosis that presents a major public health problem worldwide. The diagnosis of cancer is largely based on indicators reflecting the stage of the tumor. Various medical imaging modalities are indicated in the chest scan, which is an important diagnostic tool in 18F-FluoroDeoxy-Glucose Positron Emission Tomography with Computed Tomography (FDG PET-CT) imaging. A completely automated diagnostic system’s key means of detection is the accurate segmentation of cancer lesions in PET-CT images. The main objective of this work is to develop and test a deep learning algorithm (U-Net) for the detection of lung tumors of all stages in the FDG-PET/CT modality. The aim of this research is to develop an algorithm for the segmentation of malignant lung tumors using a convolutional neural network called U-Net. To enhance its efficiency, the transfer-learning technique is applied. The proposed network is distinguished by its efficient segmentation approach, which utilizes lightweight filtering to reduce computation and pointwise convolution to build additional features. In U-Net, skip connections were established with the “Relu” activation function to better model convergence, connecting the encoder layers to the decoder layers and allowing for the concatenation of feature maps with different resolutions. Additionally, the proposed model was trained and tested on the dataset obtained from the nuclear medicine department of the Institute of Salah Azaïez (ISA). The results showed an F-measure of 73.55%, accuracy of 99.99%, precision of 83.62%, and specificity of 99.99%. It is worth noting that the presented approach outperformed existing networks that require multiple phases of training and testing.