Background: There is an ongoing need in radiological diagnosis for more robust and objective biomarkers that correlate to patient-specific pathologies and that can show disease progression and assist in directing treatment toward optimal outcomes. Basic threshold-based densitometry, texture analysis, machine learning/deep learning techniques applied to volumetric histogram data have all been used to identify features of CT data. In the lungs features have been demonstrated to correlate with specific pathology (such as emphysema or fibrosis), physiology and clinical outcomes. Furthermore, clustering of the features has been shown to correlate with expert radiological classifications and potentially provide meaningful decision support and standardize the quantification of disease. However, many feature classification and clustering techniques are computationally expensive thereby limiting the practical usability and availability of automated quantification in real-world clinical practice.
Goal: We are developing a novel, real-time visualization mode for classification and quantification of high-resolution CT scan data that uses machine learning to enhance gray-scale image stacks by clustering and color-coding regions with similar tissue characteristics.
Methods: Pilot studies using a novel, machine learning technique developed by Boon Logic showed that coherent, near real-time voxel clustering in pulmonary CT scans is possible on a single processor core. This technique uses unsupervised learning to automatically cluster voxels with similar local tissue histogram characteristics. Our technique is extremely lightweight computationally, has low memory usage and scales well as cluster count increases. We are currently porting this technique into field-programmable gate array (FPGA) hardware to further enhance performance and allow the real-time, interactive clinical use of machine learning insights. In particular, this allows cluster templates to be created and saved to demonstrate and track disease progression in individual patients or comparison across a population of previously analyzed patients. By allowing interactive varying of the voxel neighborhood size and the percentage variation allowed within clusters, clinicians and researchers can tune the machine learning to emphasize different anatomic structures and features.
Following a validation phase through retrospective analysis of well-characterized CT data on subjects with normal lungs and a spectrum of pulmonary disease from the Lung Tissue Research Consortium project (NIIH/NHLBI publicly available database), we intend to extend this technology to CT scans of other anatomy and pathology (such as head or abdominal CT) or other medical imaging modalities, with the goal of eventually embedding the technology into clinical interpretation and analysis workflows.