Computer assisted medical intervention has become an integral part of present day’s medicine where medical image analysis plays an indispensable role. With the advancements of the modern day computing resources, machine learning techniques have emerged as a vital component in this field. The use of the supervised machine learning technique called random forests has shown very encouraging results in medical image analysis. More specifically, Random Regression Forests (RRFs), a specialization of random forests for regression, have produced the state of the art results for fully automatic multi-organ localization. Despite the very encouraging results, the relative novelty of the method in this field still raises numerous questions about how to optimize its parameters for consistent and efficient usage. Additionally, the RRF method has many parameters that require heuristic tuning which reduces its ability to be used in a more general setting. In this context, the goal of this dissertation is to carry out a detailed study on the use of the RRF methodology for multi-organ localization. First, we perform a thorough analysis of decision trees and of RRFs in the context of multi-organ localization in order to present and understand the inner workings of RRFs. From this, three directions are explored. The first direction investigates whether the localization performance of RRFs can be further improved by adding more spatially consistent information. We then propose to use the random model variables to approximate the random process. This results in a newer type of RRF, faster and more efficient in terms of memory usage : the Light Random Regression Forest. Finally, we propose an automatic and consistent approach to find the forest leaf nodes that participate in the final localization prediction. Furthermore, this proposal leads to the elimination of two other arbitrarily tuned parameters increasing the generality of RRF for multi-organ localization, without reducing their localization performances.