Towards Open processing of PACS data – TOPACS

The TOPACS project aims at large scale analysis of 3D medical images stored in hospitals. The primary goal is large scale study of the Human Anatomy (Computational anatomy). One major challenge is the size of the data to analyse, while respecting anonymity of the individuals. Keypoint extraction comes as a solution to this problem. Indeed keypoints offer a compact summary of an image, storing only important features present in the image. Each keypoint is associated with a feature vector describing the local neighborhood around the keypoint, and an efficient comparison can be computed between keypoints by measuring the distance between their respective feature vectors. During this project, we plan to be able to analyse more than 10000 individuals. The TOPACS project falls into four parts:
The first part will address keypoint extraction from 3D medical images. For 2D images, many keypoint approaches have been proposed, such as SIFT, SURF, KAZE, and recent advances in machine learning have resulted in better kepyoint algorithms, such as LIFT. But few works have proposed keypoint techniques in the field of 3D medical image processing. In this first task, we will propose new keypoint algorithms tailored to medical images, by studying both hand-crafted approaches and machine learning approaches. The proposed methods should exhibit specific characteristics : robustness to large inter-patient variability, ability to compare data extracted from different imaging modalities. We plan to extract keypoints from three hospitals, in Lyon, Saint-Etienne and Geneva.
The second part consists in devising new approaches for registration and segmentation using keypoints. A major difficulty is large scale groupwise registration. In this context, groupwise registration appears as a better means to register a large set of images, as choosing or building a single reference model would introduce a severe bias. Current approaches can register about hundreds of images together. Our goal is then to propose approaches with much larger capacities. This task will also address keypoint-based segmentation, for which few works have been proposed.
The third part will deal with statistical representations of large population as well as inference at the single-subject level. Manifold learning techniques will be considered to capture both geometric and textural normal/pathological variabilities. Classification / regression methods on manifold will then be developed to infer prediction for a given individual.

The fourth part will link the theoretical works of the first three parts to medical applications. A first task will consist in extracting data from the three hospitals, where a computer will be installed in each hospital and extract large databases of keypoints. We plan to mainly extract data from 3D CT and MRI images. A second task is the application of population analysis for anthropology, mainly for forensic science : estimating the profile of an unknown individual (gender, age, ...) or estimating the date of death. A third task will be the proposal of an online tool to provide access to the new general purpose algorithms : segmentation, registration.
More generally, the TOPACS project aims at contributing to open science, by publishing algorithms and databases, while keeping the anonymity of data present in the databases.