Modern imaging systems far exceed the human eye in spatial and spectral resolution as well as dynamic range, thus potentially allowing machine-based image pattern analysis systems to outperform manual image interpretation. In fact, recent work in pattern recognition has demonstrated that computers can equal or even surpass image classification and pattern analysis by human experts. We have previously published work investigating the progression of osteoarthritis (OA) in the human population comprising the Baltimore Longitudinal Study of Aging (BLSA). We were able to show that WND-CHARM (see AG000671-12) is able to diagnose the existence of OA in knee X-Rays with accuracies approaching that of a panel of highly trained radiologists. More recently, we have published work that WND-CHARM can predict the future onset of radiologically detectable osteoarthritis in X-Rays that were scored as radiologically clear. We were able to show that the development of moderate OA two decades in the future can be predicted with >70% accuracy from X-Rays scored as free of OA by a panel of three radiologists. Subsequently, we were able to further characterize OA progression and identify an early, slow period of change followed by rapid degeneration. We are following up our knee X-ray studies with an MRI dataset obtained from the Osteoarthritis Initiative as well as experimental MRI samples imaged here at NIA through a collaboration with Dr. Richard Spencer (NIA/LCI). In a recent study, we developed a technique using multivariate linear regression of image features derived from several types of MRI scans to construct a continuously variable cartilage quality score similar to an OARSI grade. The OARSI grade is determined histologically, and involves an invasive procedure that is not amenable to early screening or tracking disease progress. While MRI methods are non-invasive, they must first be correlated with histological grading schemes before they can be used in diagnosis or evaluating cartilage quality. We analyzed samples from cadavers that were imaged using several MRI modalities as well as scored histologically using the OARSI scale. Standard classification experiments using single MRI scans of degraded vs. non-degraded cartilage as determined by OARSI had accuracies >80%, which are substantially higher than published reports correlating scores derived from MRI images with OARSI grades. Our multivariate regression of image features from multimodal MRI scans produced a continuous score that was well correlated with the OARSI grade of the same samples (r >0.65, p <10-5). Defining a continuous grading system based on a non-invasive procedure is a key element in evaluating osteoarthritis treatment strategies. Our work with processing images of H&E stained sections of cancer biopsies has shown that consistency of sectioning and staining is a key factor in effective image analysis of medical samples. More recently we were able to demonstrate that the common approach to the use of color-sensitive image features is not as effective as deconvolving the Hematoxylin and Eosin stains into separate channels in a pre-processing step and subsequently treating them as separate grayscale channels. Additionally we analyzed many commonly used image-feature and classification algorithms to determine their relative effectiveness as well as propose strategies for improving classifier performance. Our studies of histological sections of cancer biopsies have shown that our classification accuracies are consistent with those of human readers. Our melanoma study of secondary sites of metastasis has convinced us that when the """"""""ground truth"""""""" used for training is well determined such as the known site of biopsy collection we can achieve accuracies in excess of 90% even when scoring new cases not previously used in training. We are currently searching for a comprehensive dataset composed of early or screening imagery together with definitive followup outcomes determined independently of subjective manual interpretations. While such datasets no doubt exist, they are much less common than collections consisting of single-reader assessments with no outcome data. While we can continue to process collections like these, it is becoming apparent that our classification accuracies are likely being limited by the accuracies obtainable by single human readers. Current work on studying physiological age in humans uses imaging available for several tissues through the BLSA and the Osteoarthritis Initiative (OAI). A question we are addressing is whether aging progresses through distinguishable states, as we have observed in C. elegans (see AG000674-09). We are investigating the degree to which physiological age progresses synchronously in different tissues within individuals. Additionally we are using images available through the BLSA together with physiological factors measured in study participants to look for new concurrent and predictive correlations. One of the image sets currently under study are abdominal CT scans. The viscera, bone, subcutaneous and visceral fat in these scans has been segmented into separate image masks by Dr. Sokratis Makrogiannis (NIA/LSS). Our preliminary analysis of these image masks indicates that a substantial aging signal is present in visceral fat as well as in the unsegmented whole CT scans. Even though there is generally more subcutaneous fat than visceral fat, it accounts for significantly less age-correlated signal. These preliminary results are based on cross-validation of classifiers trained on a middle-age group (56-70) and an older group (81-99). We intend to expand this analysis to broader age ranges and tissues. Another radiology project is in collaboration with Dr. Maria Knoll at the Johns Hopkins Bloomberg School of Public Health. Here the goal is to diagnose viral vs., bacterial pneumonia in children using chest X-rays. While this is not a pressing problem in the developed world, trained radiologists capable of making this diagnosis are rare in the developing world. A rapid accurate diagnosis of the nature of this disease will dramatically improve outcomes for children in developing countries. A standard set of chest X-rays is available from the World Health Organization that is well annotated, with each X-ray having been read by multiple expert radiologists forming a solid ground truth for training machine classifiers. The major challenge posed by this set is the extreme variation in the physical size of the subjects due to the variation in their ages. We have developed a strategy to compensate for this size variation by working with a student at JHSPH to manually annotate the X-rays with a set of fiducial marks that are anatomically comparable across the X-rays regardless of subject size. Using these manually aligned regions of the lung, we have very preliminary indications that such a diagnosis appears to be possible. We are currently assembling a completed manually annotated set from the WHO, as well as a different set of X-rays that has recently become available through JHSPH.
|Ashinsky, B G; Coletta, C E; Bouhrara, M et al. (2015) Machine learning classification of OARSI-scored human articular cartilage using magnetic resonance imaging. Osteoarthritis Cartilage 23:1704-12|
|Orlov, Nikita V; Eckley, D Mark; Shamir, Lior et al. (2012) Improving class separability using extended pixel planes: a comparative study. Mach Vis Appl 23:1047-1058|
|Orlov, Nikita V; Weeraratna, Ashani T; Hewitt, Stephen M et al. (2012) Automatic detection of melanoma progression by histological analysis of secondary sites. Cytometry A 81:364-73|