Enhanced Liver Image Segmentation using Custom Metrics and UNet with FastAI

Abstract

In the realm of medical image analysis, accurate segmentation of specific structures or regions is crucial for effective diagnosis and treatment planning. This Study presents an advanced methodology for segmenting medical images, specifically CT scans, using deep learning techniques within the FastAI framework. Initially, the data preparation phase involves preprocessing CT scan images and corresponding masks to create a standardized dataset. To address the challenge of significant background dominance in medical images, custom metric is proposed for foreground accuracy, which focuses on the accurate segmentation of the foreground, or regions of interest, by excluding the background from accuracy calculations. Leveraging the U Net architecture with a ResNet34 backbone, model is trained using the FastAI library, known for its high level utilities that simplify the training process without compromising flexibility. The methodology demonstrates the importance of custom metrics in medical imaging tasks and highlights the effectiveness of the U Net architecture for detailed segmentation. Our approach offers a promising avenue for medical practitioners and researchers seeking accurate and clinically relevant segmentation of medical images, ultimately contributing to better diagnostic and treatment outcomes with accuracy of 96.03%

Topics & Concepts