Resumen
Globally, lung cancer (LC) is the primary factor for the highest cancer-related mortality rate. Deep learning (DL)-based medical image analysis plays a crucial role in LC detection and diagnosis. It can identify early signs of LC using positron emission tomography (PET) and computed tomography (CT) images. However, the existing DL-based LC detection models demand substantial computational resources. Healthcare centers face challenges in handling the complexities in the model implementation. Therefore, the author aimed to build a DL-based LC detection model using PET/CT images. Effective image preprocessing and augmentation techniques were followed to overcome the noises and artifacts. A convolutional neural network (CNN) model was constructed using the DenseNet-121 model for feature extraction. The author applied deep autoencoders to minimize the feature dimensionality. The MobileNet V3-Small model was used to identify the types of LC using the features. The author applied quantization-aware training and early stopping strategies to improve the proposed LC detection accuracy with less computational power. In addition, the Adam optimization (AO) algorithm was used to fine-tune the hyper-parameters in order to reduce the training time for detecting the LC type. The Lung-PET-CT-Dx dataset was used for performance evaluation. The experimental outcome highlighted that the proposed model obtained an accuracy of 98.6 and a Cohen?s Kappa value of 95.8 with fewer parameters. The proposed model can be implemented in real-time to support radiologists and physicians in detecting LC in the earlier stages. In the future, liquid neural networks and ensemble learning techniques will be used to enhance the performance of the proposed LC detection model.