A CAD utilizing 3D massive-training ANNs for detection of flat lesions in CT colonography: preliminary results

Abstract

Our purpose was to develop a computer-aided diagnostic (CAD) scheme for detection of flat lesions (also known as superficial elevated or depressed lesions) in CT colonography (CTC), which utilized 3D massive-training artificial neural networks (MTANNs) for false-positive (FP) reduction. Our CAD scheme consisted of colon segmentation, polyp candidate detection, linear discriminant analysis, and MTANNs. To detect flat lesions, we developed a precise shape analysis in the polyp detection step to accommodate the analysis to include a flat shape. With our MTANN CAD scheme, 68% (19/28) of flat lesions, including six lesions "missed" by radiologists in a multicenter clinical trial, were detected correctly, with 10 (249/25) FPs per patient.

A spinning tangent based CAD system for detection of flat lesions in CT colonography

“…Through experimentation, we determined appropriate lower and upper bounds that retained our maximal sensitivity (82%), but also removed obvious non-lesion candidates, such as lungs and rectal tubes. We also achieved a FP rate of 31.4 per patient; these results are closer to those of the same stage (i.e., initial detection candidate detection) in a modified-shape-index-based CADe scheme [12]. Table 1 summarizes the above information and compares our results to CADe schemes based on modified shape index [12] and initial shape index.…”

“…We also achieved a FP rate of 31.4 per patient; these results are closer to those of the same stage (i.e., initial detection candidate detection) in a modified-shape-index-based CADe scheme [12]. Table 1 summarizes the above information and compares our results to CADe schemes based on modified shape index [12] and initial shape index. Our technique may be considered robust against different scanner types since the data were acquired from multiple centers and scanners.…”

“…For comparability, we used the dataset from [12]. Though a more detailed description is found there, briefly, the dataset consists of 50 CTC scans (both prone and supine for 25 patients), with a total of 28 flat lesions among them, ranging in size from 6 to 18 mm with an average size of 9 mm.…”

“…These were plusses since we did not want to risk missing a lesion due to small size or proximity to a larger mass. We performed two iterations of spatial clustering, using distance parameters based on those in [4,12] to counteract poor scan quality and the odd shapes of certain objects, such as rectal tubes and lungs. Finally, we used a simple size threshold to weed out any remaining noise and structures too large to be considered lesions.…”

“…Many of the abovementioned schemes either are not tested on non-polypoid lesions or fail to detect non-polypoid lesions while keeping the FP rate down. When Suzuki et al expanded a standard shape-index-based CADe scheme to detect flat lesions, they improved the sensitivity from 48% to 68% at an FP rate of 10 per patient by modifying the local-curvature-based shape index and utilizing massive-training artificial neural networks (MTANNs) [12]. Other recent work has targeted flat lesions [13], but there is still much to be done to make them comparable to their polypoid-detecting counterparts.…”