Prostate Segmentation

This project was done under the supervision of AI-Medic startup.
In this project I tried segmenting Transition and Peripheral Zones of Prostate.

• Here is the Link to the Dataset. (Task 5, Prostate)

Dataset Description:

• Consists of 48 MRI 3D Images with ADC and T2 modality
• 32 are shipped with their relative Mask
• Images dimensions are ~ 256 x 256 x 15
• Mask labels are as follows:
  • 0: Background
  • 1: Peripheral
  • 2: Transition

Main Challenges of the Project:

• These to Zones are almost next to each other
• Overal shape of Prostates varies significantly from case to case
• Low number of training data is troublesome specially in 3D Segmentation

2D Segmentation Workflow:

• Resizing the Images to the same resolution
• Observing the Histogram of masked zones
• Using np.clip based on the histogram
• Transforming masks to 2 separate binary groups
• Using ResNet18 Backbone pretrained on ImageNet
• Calculating the Dice score
• Updating Learning Rate using CosineAnnealingLR
• Normalizing ADC Images before and after clipping
• Implementation of CLAHE and other simillar Contrast enhancement techniques
• Using 5-fold cross validation for model evaluation

Results

Label 1:

• Overall Dices Mean = 0.77
• Overall Dices STD = 0.032
• Overall Dices Best = 0.81
• Per-Subject Dices Mean = 0.71
• Per-Subject Dices STD = 0.044
• Per-Subject Dices Best = 0.78

Label 2:

• Overall Dices Mean = 0.91
• Overall Dices STD = 0.012
• Overall Dices Best = 0.92
• Per-Subject Dices Mean = 0.86
• Per-Subject Dices STD = 0.023
• Per-Subject Dices Best = 0.90