MICCAI 2013 Highlights and Impressions

From Sept. 22 to Sept. 26, this year’s MICCAI was held in Nagoya, Japan. Stephen Aylward and myself (Roland Kwitt) participated in the conference and presented our paper Studying Cerebral Vasculature Using Structure Proximity and Graph Kernels.

Since Stephen’s flight was cancelled, we were traveling together with Marc Niethammer and Martin Styner of UNC from Detroit to Nagoya. In fact, it seemed as if half of MICCIA’s Program Committee was on that flight! This raised the funny question if travel plans should be more coordinated to have greater diversity, since you obviously don’t want to jeopardize the whole community if something would happen to that plane 🙂 Fortunately, everything went well and we arrived ~13 hours later in Nagoya, Japan. While North Carolina is definitely an awesomely hot place, Nagoya seemed to to top that, mostly because there wasn’t an abundance of air conditioning. 

Some Statistics

MICCAI received 798 submissions this year with 262 papers being accepted (32%), 37 (4.6%) as oral and 225 as poster presentations (28.1%). This compares to 781 submissions and 252 accepted papers at last year’s MICCAI in Nice, France.

Presentations & Posters

In terms of interesting presented papers, below is a list of my personal highlights (although, theses are just few, out of many interesting papers this year). In no particular order:

W. Wein et al., Global Registration of Ultrasound to MRI Using the LC^2 Metric for Enabling Neurosurgical Guidance

As the title already says, the main theme was to establish a suitable similarity metric for registering US images to MR scans. This was a particularly interesting presentation, since cross-modality registration seems to become one of the more dominant themes these days (again).

M.K. Chung et al., A Persistent Homological Sparse Network Approachto Detecting White Matter Abnormality in Maltreated Children: MRIand DTI Multimodal Study

Here, the authors apply ideas from the field of computational topology to overcome the problem of the inherently univariate (in the response variable) tensor-based morphometry approaches to study population differences in MR and DTI. This was interesting work, primarily since concepts from computational topology are swapping over to the vision and medical community more and more …

T. Brosch & R. Tam, Manifold Learning of Brain MRIs by Deep Learning

In this work, the authors propose an approach that uses Deep Belief Networks (DBN) to learn the manifold of MRI images. While their results in correlating the manifold coordinates with disease parameters (e.g., Alzheimers) are mediocre, the advantage is that there is no longer a need to define a similarity measure. This “first shot” at the problem demonstrates that deep learning definitely has potential even in the medical field.

D.H. Ye et al., FLOOR: Fusing Locally Optimal Registrations

This was an interesting registration work, in which the authors dismiss the idea of one globally optimal registration solution and instead try to fuse candidate registrations, where the candidates are generated by varying the image domain and the smoothness parameters of a given registration algorithm (e.g., Demons).

Another incarnation of random forests (for the 4th year in a row now 🙂 won one of the “MICCAIYoung Scientist Awards”:

D. Zikic et al., Atlas Encoding by Randomized Forests for Efficient Label Propagation

The authors introduce a new approach to multi-atlas label propagation where, typically, each atlas (image + labelmap) is registered to the target. In their approach, each individual atlas is encoded in reference to one probabilistic atlas. This avoids multiple time-consuming registrations at test time, however, it also reduces robustness. Nonetheless, the paper shows competitive results to the state-of-the-art at a fraction of the runtime required for traditional methods.

The last paper I am gonna mention is by

T. Heinmann et al., Learning without Labeling: Domain Adaption for Ultrasound Transducer Localization

Here, the objective was to effectively train a detector for a TEE transducer in X-ray images without having to go through the effort of labeling a lot of images. Their idea is simple, yet very effective:Through simulation of the TEE transducer appearance in X-ray (from a CT volume), an arbitrary amount of training data can be generated to train a discriminant classifier. Yet, the training and target domains are different. To cope with that, the authors use ideas from domain adaption (a common theme in computer vision recently).

The Future of MICCAI

At the end of the conference, there was a panel discussion on “The future of MICCAI”, led by Nassir Navab and several established members of the community (N. Ayache, J. Duncan, etc.). A common theme, pointed out by several people, was to look for a closer integration with clinical practice in order to avoid solving the wrong problems (or problems which are actually no problems at all).

Another interesting discussion was on “information transfer”: Most algorithms will produce some reasonable results (hopefully); These results might be useful to the scientific community, however, little is done in terms of presenting those results in a clinically relevant way to the physicians.

Summary

In summary, there was a decent amount of good papers this year. A dominant theme (and I might be biased here) was to use more concepts from the machine learning community (manifold learning, discriminative classifiers, subspace identification, etc.) to tackle traditional problems from a different angle.