Tissue Classification and Disease-Specific Imaging Based on Machine Learning

Jihye Baek

Supervised by Kevin J. Parker

601 Computer Studies Building

Tissue classification has shown promise for enhancing the efficacy of medical imaging modalities, which is utilized for diagnosing disease, monitoring disease progression, and tracking treatment response. For classification, a number of quantitative parameters extracted from medical images have been developed and employed for computer-aided diagnosis with artificial intelligence (AI).

This dissertation proposed an approach to best integrate multiple parameters, including feature extraction from raw radio frequency (RF) data, in a straightforward imaging framework utilizing support vector machine (SVM). Though several medical imaging modalities could be used to validate this approach, ultrasound imaging was employed due to its speed, wide availability, and relatively low cost. First, to improve machine learning classification for diagnosis, the H-scan analysis was used to extract information from RF data to estimate frequency components. The H-scan can detect subtle changes in tissues and attenuation coefficient, and thus the measures using the H-scan contributed to more accurate classification by adding frequency-dependent information. This thesis fine-tuned the H-scan for more precise parameter estimation. Its performance was verified by (1) applying it to phantoms, livers, kidneys, and melanoma cases and (2) comparing it with bioluminescent imaging (BLI) as a reference and shear wave elastography (SWE). The H-scan showed a strong correlation with BLI and outperformed SWE in tracking the progression of pancreatic cancer metastasis in the liver. In addition to the H-scan features, B-mode texture patterns, histogram-based speckle parameters, SWE measures, and morphological features were extracted, but employed features were selected depending on target organs and diseases to be detected.

To assess the accuracy of the features to characterize tissue signatures, SVM was utilized to classify liver conditions, including normal, inflammation, progressive steatosis, and fibrosis. The resulting classification accuracies ranged from 94.6 to 100% and, further, the contribution from the two H-scan parameters, estimating scatterer size and attenuation, was found to be superior to other features. The constructed SVM hyperplanes can be used for disease diagnosis.

Parameter trajectories were investigated as diseases progressed as means of predicting disease severity in addition to disease diagnosis. We investigated liver steatosis, fibrosis, and pancreatic cancer metastasis in the liver, and each disease followed a specific trajectory as it progressed over time. Moreover, we suggested the disease-specific imaging (DSI) framework to visualize the trajectories with SVM classification. SVM assigns a specific color for each disease, and the projection of a measured feature onto a specific trajectory curve quantifies the color intensity of DSI. A trained DSI was assessed by applying it to independently acquired steatosis data, which showed that DSI images illustrated progressive steatosis better than conventional B-mode images and comparable to histology im- ages. Moreover, another DSI was trained with 7 parameters, including the H-scan and SWE parameters. Its performance was compared with magnetic resonance imaging (MRI)-proton density fat fraction (PDFF), which demonstrated that the DSI outperformed MRI-PDFF quantification and imaging.

Finally, this thesis modified the DSI to achieve higher performance, and suggested a method to estimate the probability of disease severity utilizing the distance between the SVM hyperplane and each feature in 3D space. This new DSI was applied to an in vivo human breast study. Its performance of breast lesion detection was compared with that of radiologists and a commercially available deep learning system. The DSI yielded the highest area under the curve (AUC), accuracy, sensitivity, and specificity. Furthermore, DSI color-coded the probability of malignancy with a color bar, providing corresponding Breast Imaging Reporting and Data System (BI-RADS) scores.

In summary, this dissertation developed the DSI, incorporating frequency- dependent features from RF data into machine learning, and demonstrated the feasibility of applying it to medical ultrasound imaging. Furthermore, while this dissertation investigated only ultrasound signals, the proposed framework has the potential to be applied to other imaging modalities that can provide quantitative multiparametric features.