Abstract:
Liver cancer remains a critical global health concern, where early and precise diagnosis
significantly impacts treatment effectiveness. Traditional methods for liver tumor segmentation
and classification often suffer from issues such as image noise, unpredictable tumor morphology,
and time-consuming manual processes, making them inefficient for clinical applications. This
research aims to overcome these limitations by developing a robust deep learning-based system
for automated liver tumor segmentation and classification using 2D CT images.
The project leverages a U-Net model with a MobileNetV2 backbone for segmentation and a
custom CNN architecture for classification, distinguishing between malignant and benign tumors.
Data preprocessing techniques, including normalization, augmentation, and noise reduction, have
been employed to enhance image quality and improve model performance. The model is trained
on manually labeled LiTS CT scan datasets, labeled in collaboration with a radiologist, addressing
the lack of publicly available labeled data.
The initial segmentation results achieved a Dice score of 0.75–0.78 and an IoU of 0.45–0.50,
demonstrating promising segmentation performance. The classification model achieved 93%
validation accuracy, highlighting the need for further robustness testing on external datasets.
Segmentation results received 95% of accuracy.
Classification results received 93% of accuracy.
