抽象的な

Computer-aided diagnosis in digital mammography: comparison of two commercial systems

Donato Cascio, Francesco Fauci, Marius Iacomi, Giuseppe Raso, Rosario Magro, Debora Castrogiovanni, Guido Filosto, Raffaele Ienzi & Maria Simone Vasile

Aim: Within this work, a comparative analysis of two commercial computer-aided detection or diagnosis (CAD) systems, CyclopusCAD® mammo (v. 6.0) produced by CyclopusCAD Ltd (Palermo, Italy) and SecondLook® (v. 6.1C) produced by iCAD Inc. (OH, USA) is performed by evaluating the results of both systems application on an unique set of mammographic digital images routinely acquired in a hospital structure.
Materials & methods: The two CAD systems have been separately applied on a sample set of 126 mammographic digital cases, having been independently diagnosed by two senior radiologists. According to the human diagnosis, the cases in the sample reference set are divided into 61 negatives and 65 pathological cases (21 cases displaying both mass lesions and microcalcifications and 44 cases characterized only by mass lesions). The images in the pathological subset contain 123 human diagnosed mass lesions and 37 human diagnosed microcalcifications clusters. In the case of CyclopusCAD, the system offered the possibility to evaluate sensitivity at several threshold levels (working points); five different setting levels (high sensitivity, normal sensitivity, standard, normal specificity and high specificity) have been used. Results: At the standard threshold level, CyclopusCAD exhibits an overall sensitivity of 83.1 versus 66.2% for iCAD (p = 0.04) and an average number of false positives per image (FP/im) of 1.38 against 0.47 for iCAD (p < 0.01). Specifically, for the mass lesions, CyclopusCAD exhibits a sensitivity of 76.9% at a rate of 0.73 FP/im, while iCAD displays a sensitivity of 61.5% at 0.28 FP/im. For the microcalcifications, CyclopusCAD exhibits a sensitivity of 76.2% at a rate of 0.64 FP/image, while iCAD displays a sensitivity of 61.9% at 0.19 FP/im. The reported results have also been expressed in terms of free-response receiver operating characteristic curves, corresponding to five different thresholds in the case of CyclopusCAD and to one single threshold value for iCAD. Conclusion: The overall accuracies of the two systems are fairly comparable up to the uncertainty level of this analysis. CyclopusCAD may reach a higher sensitivity level for both masses and microcalcifications owing to the flexibility in the working point choice, with the price of a major number of FP/im.