User: Guest  Login
Document type:
Wiestler, B.; Baur, C.; Eichinger, P.; hwiestler; Zhang, T.; Biberacher, V.; Zimmer, C.; muehlau; Kirschke, J.; Albarqouni, S.
Fully Automated Multiple Sclerosis lesion detection on multi-channel subtraction images through an integrated Computer Vision- Machine Learning pipeline
Background: MR imaging is a mainstay of MS disease activity as- sessment. However, lesion detection is time-consuming and subject to relevant inter-observer variability. We have therefore aimed to develop an integrated analysis pipeline for automatic detection and segmenta- tion of new or enlarged MS lesions in subtraction images over time. Methods: Initially, a multi-step computer vision pipeline marks can- didate lesions on both DIR & T1 subtraction images, using constraints on shape and intensity. Next, shape, intensity and neighborhood fea- tures were extracted from each lesion, and an oblique random forest (RF) classifier was trained on a set of segmented subtraction maps of 42 patients. Classification accuracy was validated in 22 previously un- seen patients. Results: In the validation cohort, the computer vision pipeline marked 866 candidate lesions. Of these, 207 were true lesions, while 8 true le- sions were missed, mostly because of their small size (detection rate 96.2%). The RF classifier for separating noise from true lesions reached a total accuracy of 95.3% (Fig. 1A), detecting 192 of 207 marked lesions (92.7%, Fig. 1B). Intriguingly, the algorithm also identified nine MS lesions which have previously been overlooked by two neuroradiologists. The software therefore correctly labelled 192/215 new lesions (detection rate 89.3%). Conclusions: A Computer Vision-Machine Learning pipeline rely- ing on shape, intensity and neighborhood features can detect new or enlarged MS lesions with a compelling detection rate of 89%, which is well within the range of human readers in contemporary trials. Intriguingly, the pipeline even detected previously overlooked MS lesions, highlighting the potential of computer-assisted radiology.
Journal title:
Clinical Neuroradiology
Journal volume:
Journal issue:
Pages contribution: