Locally Adaptive Probabilistic Models for Global Segmentation of Pathological OCT Scans