1Nakajima Laboratory Department of Electrical Engineering Keio University, 2Institute of Information Science and Electronics University of Tsukuba, 3Department of Obstetrics and Gynecology Keio University School of Medicine, 4Department of Clinical Research National Ohkura Hospital
We attempted to develop an accurate, automatic method for determining the volume of internal organs.
Our method has three steps. In the first step, three-dimensional sonographic data are obtained from an organ. In the second step, organ boundaries are accurately defined. For this purpose, we designed a neural
network system with which we classified each voxel as liquid, soft tissue, or boundary. Because this system
alone may not always be accurate, however, we devised step three, the ‘bubble filling methods.’ With the
combination of the neural network system and the bubble filling method, we were able to determine the area
of the target organ. Volume was obtained by calculating the number of voxels inside the determined area.
To evaluate the accuracy of this method, we prepared seven phantoms of different volumes, ranging from
0.31 to 39 ml, and compared volumes computed by this method with these actual values. Mean difference
between actual and computed volume was 0.729 ml (3.38%). We then computed the volumes obtained with
another method, in which we estimated the boundaries on visual inspection of three-dimensional sonograms
(manual method). Mean difference between actual volume and volume estimated by the manual method
was 0.641 ml. We treated volumes obtained with the manual method as standard values, because actual organ volume was unknown. Fifty three-dimensional fetal bladder images were obtained from eight fetuses.
We used this method to compute fetal bladder volume (y) and compared the values obtained with those
obtained with the manual method (x). The linear regression line of y=-0.0741+1.02x indicates that this
method may be sufficiently accurate to use in calculating the volume of the fetal bladder.