Liver Tumor Segmentation with Deep Learning Architectures
Develop an automated segmentation model for liver tumors in CT scans and compare the effectiveness of U-Net and ResNet50 architectures in handling irregular boundaries and class imbalance.
🧠 Domain: Medical Imaging & Deep Learning
🎯 Task: Segmentation (Tumor vs. Background)
📂 Dataset: LiTS – Liver Tumor Segmentation Dataset (130 CT scans)
Project Domain
Imaging
Task
Segmentation
The Goal:
How U-Net and ResNet50 Architectures assist in accurate and efficient medical imaging?
1
Methodolgy:
Implemented U-Net with skip connections for precise boundary capture.
Fine-tuned ResNet50 (pre-trained on ImageNet) as an encoder with a custom decoder for segmentation.
Used the LiTS dataset (130 CT scans) with preprocessing: normalization, resizing (128×128), and augmentation (rotations, flips, intensity scaling).
Applied Tversky loss (α=0.7, β=0.3) to handle class imbalance.
Training setup: Adam optimizer, LR=1e-4, ReduceLROnPlateau scheduler, batch size=16, 20 epochs.
Post-processing with morphological operations to refine segmentation masks.
The Result (Outcome / Conclusion)
U-Net: Dice = 0.93, Jaccard = 0.92, Accuracy = 97.18%
ResNet50: Dice = 0.94, Jaccard = 0.94, Accuracy = 97.15%
ResNet50 achieved slightly better segmentation for small/irregular tumors but was computationally heavier.
The study demonstrated that both architectures are effective: U-Net is faster and lightweight, while ResNet50 excels at capturing finer details.
2
The Result
U-Net: Dice = 0.93, Jaccard = 0.92, Accuracy = 97.18%
ResNet50: Dice = 0.94, Jaccard = 0.94, Accuracy = 97.15%
ResNet50 achieved slightly better segmentation for small/irregular tumors but was computationally heavier.
The study demonstrated that both architectures are effective: U-Net is faster and lightweight, while ResNet50 excels at capturing finer details.
3
