Artificial Systems Can Complement Human Vision in Medical Imaging

Texture analysis is an emerging field; and it is just beginning to integrate with radiology. Carrying out research with thousands of images can be overwhelming, without an effective and efficient sorting algorithm. The aim of this experiment was to develop a sample selection-elimination protocol for a large research project seeking to compare fetal 1.5- tesla versus 3-tesla magnetic resonance images. Firstly, we had to find optimal methods for image selection. In a compiled database of 1.5-tesla and 3-tesla images, we began by manually selecting sequences based on discernible-anatomical structures (ventricle, thalamus, grey matter, white matter). Then 1.5-tesla and 3-tesla image batches were categorized into two groups based on gestational age (i.e. first group: 20-28 week; second group: 29 week). The final stage was sample elimination by variance and by real bit-depth – that is the actual stored bits which correlate with the measured mean, not the expected mean derived from allocated bits reported in the image metadata. Though both 1.5-tesla and 3-tesla images were fo rma t t ed a s 16-bit digital imaging and communications in medicine, a maximum o f 12 and 16 stored bits were measured in 1.5-tesla and 3-tesla respectively. This finding was crucial for fair selection and therefore ideal for assessment of the quality of 1.5-tesla versus 3-tesla – rather than simply relying on visual appearance. Stored bits were constant in both gestational groups of the same magneti