Lung Tumor Image Segmentation from Computer Tomography Images Using MobileNetV2 and Transfer Learning

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Oncology
• Deep Learning for Medical Diagnosis

Background:

• Lung cancer remains a leading cause of cancer mortality globally.
• Early detection is crucial, but symptoms often appear in advanced stages.
• Accurate segmentation of cancerous lesions in CT images is vital for automated diagnosis.

Purpose of the Study:

• To develop an improved hybrid neural network for semantic segmentation of malignant lung tumors in CT images.
• To enhance the accuracy and efficiency of automated lung cancer detection systems.

Main Methods:

• A hybrid neural network fusing MobileNetV2 and UNET architectures was developed.
• Transfer learning using pre-trained MobileNetV2 as an encoder for feature extraction.
• Implementation of lightweight filtering, pointwise convolution, and skip connections with Relu activation for improved convergence and feature map concatenation.

Main Results:

• The proposed network achieved a Dice score of 0.8793, recall of 0.8602, and precision of 0.93 on a validation dataset.
• The hybrid model demonstrated superior performance compared to existing segmentation networks.
• The approach effectively reduces computation while enhancing feature building.

Conclusions:

• The developed hybrid neural network offers an efficient and accurate solution for lung tumor segmentation in CT images.
• This advancement contributes to the development of fully automated diagnostic systems for lung cancer.
• The model's performance indicates its potential for clinical application in early lung cancer detection.