Unsupervised Learning of Finite Mixture Models
IEEE Transactions on Pattern Analysis and Machine Intelligence
Laplacian Eigenmaps for dimensionality reduction and data representation
Neural Computation
Face Recognition Using Laplacianfaces
IEEE Transactions on Pattern Analysis and Machine Intelligence
Optical Biopsy Mapping for Minimally Invasive Cancer Screening
MICCAI '09 Proceedings of the 12th International Conference on Medical Image Computing and Computer-Assisted Intervention: Part I
MICCAI '09 Proceedings of the 12th International Conference on Medical Image Computing and Computer-Assisted Intervention: Part I
Probabilistic Region Matching in Narrow-Band Endoscopy for Targeted Optical Biopsy
MICCAI '09 Proceedings of the 12th International Conference on Medical Image Computing and Computer-Assisted Intervention: Part I
MICCAI'10 Proceedings of the 13th international conference on Medical image computing and computer-assisted intervention: Part II
MICCAI'10 Proceedings of the 13th international conference on Medical image computing and computer-assisted intervention: Part II
Hi-index | 0.00 |
Recent introduction of probe-based confocal laser endomicroscopy (pCLE) allowed for the acquisition of in-vivo optical biopsies during the endoscopic examination without removing any tissue sample. The non-invasive nature of the optical biopsies makes the re-targeting of previous biopsy sites in surveillance examinations difficult due to the absence of scars or surface landmarks. In this work, we introduce a new method for recognition of optical biopsy scenes of the diagnosis endoscopy during serial surveillance examinations. To this end, together with our clinical partners, we propose a new workflow involving two-run surveillance endoscopies to reduce the ill-posedness of the task. In the first run, the endoscope is guided from the mouth to the z-line (junction from the oesophagus to the stomach). Our method relies on clustering the frames of the diagnosis and the first run surveillance (S1) endoscopy into several scenes and establishing cluster correspondences accross these videos. During the second run surveillance (S2), the scene recognition is performed in real-time and in-vivo based on the cluster correspondences. Detailed experimental results demonstrate the feasibility of the proposed approach with 89.75% recall and 80.91% precision on 3 patient datasets.