The Traveling Salesman Problem with Neighborhoods (TSPN) generalizes the classical Traveling Salesman Problem (TSP) by requiring a tour to visit a set of polygonal regions rather than fixed points, a natural goal that arises in various applications. While the geometric TSP allows arbitrarily close approximation and provably optimal solutions for benchmark instances of significant size, the TSPN is considerably more challenging, both in theory (due to APX-hardness) and practice, for which only benchmark instances up to 16 regions have been solved to optimality. Here we present a branch-and-bound algorithm that combines a spectrum of geometry-based filters (for reducing the number of considered sequences) with Second-Order Cone Programs (SOCP) (for computing optimal tours for a given permutation of neighborhoods). This allows us to solve larger polygonal TSPN instances than before to within an optimality tolerance of 0.1%; moreover, while previous work (both in theory and practice) relied on relatively benign neighborhoods, we can handle non-convex, non-simple neighborhoods of different sizes. In experiments on 490 benchmark instances with up to 50 polygons each, our method achieves a 99.6% optimality rate within 300s, with the remaining two instances solved within 595s. For 68 larger instances of size n = 60, our method still allows solving 86.8% of instances to optimality within 900s, leaving only 3 of the instances with optimality gaps above 3%, with the maximum being 5.53%.