Postdoc – Deep learning for colon cancer histopathological images analysis

The Engineering science, computer science and imaging laboratory (ICube, Strasbourg, France), associated with the Institute of Research in Computer Science, Mathematics, Automatics and Signal processing (IRIMAS, Mulhouse, France), opens a postdoctoral position for a computer scientist, in the field of artificial intelligence and histopathological whole slide images analysis, with a duration of 34 months (2020/02/01 – 2022/11/30).

In the context of the AiCOLO project described above, the appointee will work in close collaboration with the three partners of the project to develop the methodology for WSI analysis, spatial patterns extraction and the machine learning approach to classify automatically the genetic mutation from HES images.

More specifically, the objective is to develop a complete methodology enabling to assign a label to each image region. This problem will be tackled by two complementary approaches: a pixel-based method, in order to obtain a cartography of regions of interest for the studied pathologies (inflammatory zones or tumors area) and an object-based approach enabling to compute a list of biological objects with their contour, localization and attributes. The main workflow must be automatic in order to operate in an unsupervised manner, which constitutes a crucial and challenging aspect of this task. The WSI analysis system will rely both on previous work2, 3 and on novel techniques based on levellines decomposition of an image4 and connected operators from mathematical morphology5 based on hierarchical representations. These latter methods enable to analyze an image at the level of the connected components of its threshold sets (or other increasing transformation).

These methods are relevant in this context since (i) they enable to process an image in a contrast invariant way; (ii) they prevent to alter the contour of objects; (iii) they permit to compute object based attributes. The work will also consists on the identification of genetic prognostic/predictive markers on HES slide. BRAF and RAS mutational status are mandatory required for the treatment of metastatic colon cancer. These markers are both prognostic and predictive of response to anti EGFR therapies. Recent data in lung cancer make the demonstration that gene mutations could affect the pattern of tumor cells on a lung cancer whole-slide image6. Training network using the presence or absence of mutated genes as a label revealed that there are certain genes whose mutational status can be predicted from image data alone: EGFR, STK11, FAT1, SETBP1, KRAS, and TP53 with good accuracy. The ability to quickly and inexpensively predict both the type of cancer and the gene mutations from histopathology images could be beneficial to the treatment of patients with cancer given the importance and impact of mutation in patient care. We propose to perform similar work on colon cancer and try to isolate feature detected by neural network to detect these mutation.

This research will be part of a collaboration between AI researchers of ICube and IRIMAS, and pathologists and biostatisticians of the Centre Georges Francois Leclerc, Cancer and adaptive immune response team (Dijon). Regular exchanges will take place either by videoconference or during stays in their structures, to benefit from their expertise. Within the ICube laboratory, the postdoctoral researcher will be integrated into the Data and Knowledge Science team under the supervision of Cédric Wemmert and Benoît Naegel, and co-supervised by Germain Forestier and Jonathan Weber (IRIMAS).