Year: 2024
Type: conference-paper
Abstract: Lung cancer has the highest rate of cancer-caused deaths, and early-stage diagnosis could increase the survival rate. Lung nodules are one of the most common lung cancer signs, therefore, the development of lung nodule detection systems becomes significantly crucial. Different lung nodule detection models have been introduced, however, they haven't fully considered all aspects of efficiency. Hence, in this paper, we take a more efficient approach in introducing a lung nodule detection model while providing improvements in nodule detection. In this regard, we consider 2D slices of the CT scans which not only lead to lower computational load and complexity in training and inference phases but also, allow us to examine the potential of 2D models in lung nodule detection. Moreover, we utilize the tiny version of Swin Transformer to leverage the benefits of Vision Transformers (ViT) while having a much lower computational complexity. We also build a Feature Pyramid Network on top of our backbone to detect nodules of various sizes better, especially small nodules. Additionally, we consider Transfer Learning for the training procedure, resulting in significantly lower training iterations. The experimental results demonstrate the competitive performance of our model compared to the state-of-the-art models with even higher mAP and mAR for small nodules by 1.3% and 1.6% respectively. Also, our proposed model achieved the highest mAP in all size nodules with 94.7% and mAR of 94.9%.
Authors: Saeed Shakuri, Alireza Rezvanian
Institution: University of Science and Culture
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4407785605