Analysis of liver lesions in dynamic contrast enhanced MR images

Lesion detection and characterization is an important step in the diagnosis and treatment planning of focal liver lesions such as hemangiomas, adenomas, and liver tumors. Image processing and image analysis using machine learning can aid radiologists performing these tasks. In this thesis several methods regarding image registration, lesion detection, and lesion classification were developed and evaluated. The methods were developed for dynamic contrast enhanced (DCE) MR images, sometimes in combination with other MR sequences.
DCE-MR images contain both structural and functional information, which makes them very useful for image analysis. However, motion between the different contrast enhancement phases is introduced by inconsistent breath hold and by bowel and cardiac movements. We evaluated two image registration methods, a groupwise and a pairwise registration method. The results showed that the groupwise registration achieved better temporal alignment with smoother spatial deformations than pairwise registration. The results provided by the groupwise registration are such a satisfactory that clinical abdominal DCE-MR images are automatically registered and presented to the radiologists for analysis in their current workflow. Moreover, the groupwise registration method was used for the other studies described in this thesis.
Next, a liver segmentation method using convolutional neural networks (CNN) which exploits DCE-MR images was developed. The liver segmentation was used as a first step in the liver lesion detection method, to define the region of interest. The second step was the introduction of a novel liver lesion detection CNN method. We found that DCE-MR images alone were not sensitive enough for the liver lesion detection method. The inclusion of diffusion weighted (DW) MR images was proven to be a valuable addition to the detection method, since the detection rate increased with more than 50 per cent.
The described lesion detection method is a general method and could fail if a patient presents features not known or seen before by the CNN. We proposed a patient-specific fine-tuning approach to obtain a detection method dedicated towards one patient. During patient-specific fine-tuning, previously acquired MR examinations were exploited to update the pre-trained CNN towards the patient-specific features. This approach improves the performance of lesion quantification methods.
Once the lesions are detected, the type of lesion needs to be determined. We proposed lesion classification method, which was able to successfully distinguish benign from malignant lesions as well as differentiating five focal liver lesions: adenoma, cyst, hemangioma, HCC, and liver metastasis. Contrast curve, gray level histogram, and gray level co-occurrence matrix texture features were extracted from the DCE-MR and T2-weighted MR images. In addition, risk factors including the presence of steatosis, cirrhosis, and a known primary tumor were used as features. The selected features originated from all four feature categories and from DCE-MR and T2-weighted MR images, which underpins the importance of a wide selection of features.