A study of deep learning for automatic segmentation of healthy liver in abdominal computed tomography scans

Medical-image segmentation refers to a process in which regions of interest (ROIs) such as organs are annotated in 2D or 3D images. Medical-image interpretation performed by radiologists and physicians has proved to be crucial for early clinical detection, diagnosis, and treatment. However, precise manual segmentation of medical images is a time-consuming process, which is an issue when reducing the time gap between medical scanning and any required medical procedure would be crucial.

In view of the above, the use of computers to assist medical-image interpretation, in the form of computer-aided diagnosis (CAD), has become an essential tool for radiologists. In recent years, automatic image segmentation based on deep learning (DL) models has become popular due to the fast and precise results that could be achieved, thus surpassing traditional methods.

Although DL models have outdone traditional methods, the segmentation required in the medical sector requires precision that DL models are yet to achieve. The high variability from patient to patient, ROI overlapping, limited size of datasets to learn from, and low-resolution images, are the main factors hindering the development of a universal DL model suitable for any specific problem.

The first stage of this project was a review of existing solutions, architectures and implementations for medical-image segmentation, highlighting key concepts that emerged. Subsequently, state-of-the-art deep neural networks (DNNs) were implemented, namely: U-Net and 3D U-Net. A proposed model was then implemented, where medical images were first preprocessed and classified to include the liver before passing through the DNN.

The models were applied and compared to the liver CT scan dataset publicly provided by the CHAOS challenge, where the CT scans were acquired at portal phase after contrast agent injection for pre- evaluation of living liver donors. The proposed model was found to improve over the precision of U-Net, from a DICE score of 0.94 to 0.97.

Figure 2: Proposed Model

Student: Jeremy James Cachia
Course: B.Sc. (Hons.) Computer Engineering
Supervisor: Prof. Johann Briffa