PMC

The fraction of lesions correctly localized, plotted as a function of radius threshold for correct localization, for each reconstruction algorithm and the CNPW observer.

CNPW results for all algorithms showing the dependence of A_LROC upon the number of iterations. For each datapoint, the filter which maximized A_LROC was used. The LOR-OSEM3D algorithms showed continued improvement out to the later iterations studied, whereas the other algorithms experienced slight losses in performance at the later iterations.

Comparison of P_LOC and A_LROC results for the CNPW observer versus human observers. The CNPW results shown here are for the same 110 test images as read by the human observers; however, strict task-equivalence did not apply since the human observers were permitted to adjust the image display greyscale whereas no such ability was built into the CNPW observer.

The fraction of lesions correctly localized by the human observers, plotted for each reconstruction algorithm as a function of radius of correct localization. The data show the classic pattern of a fast initial rise, a plateau (where the radius threshold was selected), followed by a gradual rise due to random selection of neighboring sites.

Example CT images of the phantom chest (top row), abdomen (bottom left), and pelvis (bottom right). The chest images show a lesion in the left lung (left) and mediastinum (right). The structure in the lungs arises from the packing of Styrofoam beads in nylon bags, and the structure in the mediastinum, abdomen, and pelvis body compartments arises from small air bubbles suspended in the open-cell foam material.

The image display interface used for the human observer studies presented a single 2D image. The observer was asked to select a confidence rating by clicking on the six-point scale as shown, and mark the location in the image most likely to contain a lesion (red crosshairs). When training, the interface then would provide feedback as to the truth (lesion present versus absent) and actual lesion location (when present).

CNPW results, showing how A_LROC changed as a function of filter width (s.d.) from 0.0–2.0 voxels, with a different line for iterations 1–10. The parameters providing the highest A_LROC for each algorithm are shown on the plots. In each case, the difference in performance between successive iterations got smaller as the number of iterations increased, and the difference in A_LROC between 6 and 10 iterations was relatively small for all algorithms.

Example maximum intensity projection (MIP) images of the scans with lesion present on day 1 (left) and day 2 (right). These images were formed by averaging the images for all four scans from each day to obtain low noise images so that the lesions could be easily visualized, and the upper limit of the greyscale was lowered to enhance visualization of the body compartments.

The whole-body lesion detection phantom consisted of a brain phantom, thorax with liver and lung compartments, and elliptical pelvis with bladder compartment. The lungs were filled with nylon mesh bags of Styrofoam beads, and the body and pelvis compartments were filled with a low water resistance open cell foam. ⁶⁸Ge-infused silicone lesions were mounted in the mediastinum, lungs, abdomen, liver, and pelvis compartments. The close-up at the bottom shows an 8-mm lesion placed in the lung (black arrow), and a 12-mm lesion inserted into the open-cell foam of the body compartment (white arrow).

Example transaxial slices, each containing a lesion, for the four reconstruction schemes with six iterations and filters shown in : (from top to bottom) FORE-OSEM2D, AW-OSEM3D, LOR-OSEM3D, and LOR-OSEM3D+PSF. The images show a sampling of phantom slices ordered from more superior (left) to more inferior (right) axial locations. Note that identical scan data were reconstructed for each case, and all differences are due to the different reconstruction processing. Striking differences in noise texture and depiction of the lesions can be seen for the different reconstruction algorithms.