CMIR Hackathon ICC 2020
Our project relies heavily on machine learning including deep learningand transfer-learning. Studying medical informatics and medical enge-neering has brough our team a wide range of experiance in the healthAI field. We got inspired by the talk from Prof. Dr. Klaus Maier-Hein, DKFZ Heidelbreg about Deep learning in medical imaging“ at theHeiKa Symposium Artificial Intelligence in medicine 2019 - effects andside effects. In the state of the art review from Litjens et al propose that”advances in the computer vision community with respect to caption gen-eration in natural images, combining text and image analysis with RNNsand CNNs, will likely soon make their way into medical image analysis.[..] We expect that more tasks-specific architectures will start appearingin the coming years as well, for example in registration and Content BasedImage Retrieval”. The Content Based Image Retrieval is a task thatcan be applied to a lot of data sets including PACS. We will create anecosystem for information retrieval for several diseases. Our idea is goodfitted for a hackathon as it is scalable in it’s size. The hyperparamters,data size and diseases can be tested until the end of the hackathon.
Read more about Content Based Image Retrieval and finalize the concept. Learnas much as possible about possible diseases and available datasets. Try toevaluate those and make an order of the tasks.
Design and test scripts
Jupyter-notebooks are written and tested for possible diseases. This stage isvery variable in time as the amount of diseases can be increased if there is muchmore time. This stage can be done until the final day of the hackathon.
Finalize & GUI
Bring all notebooks together in one suite and finalize the software. This meansthat a user friendly GUI is put together that can be tested.
Presentation of the results
Do the final presentation of our work. This includes a proposal and life demoof the wide range of features that we implemented.
Image reconstruction including Content Based Image Retrieval cound be asnew type of deep learning models according to Litjens et al.. These systemscan potentially useful in every aspect of medicine where images are used for adiagnosis. The error rate of these new algorithms can be much lower comparedto human level in various task. Already existing feature based systems couldsimply be extended with more features.
With our project we hope to expand the state of the art in medical image anal-ysis a little further. We know of the hardships when dealing with unregisteredand unspecified images in the context of AI. We hope to create a prototype thatis not bound by specific anatomical setting but is universally applicable.The Content Based Image Retrieval is a task that be applied to a very wide rangeof medical data sets including Picture Archiving and Communication System(PACS). We will create an ecosystem for information retrieval for several dis-eases. During the hackathon we will try to find as much medical imagining datasets as possible and then start with the most promising ones. Using does imageswe will apply the Content Based Image Retrieval. What type of information isneeded to be retrieved depends heavily on the context of the specific disease.The information can be very simple, for example the prop ability for cancerusing an RGB-image of a melanoma. But it can be also much more complexby extracting multiple features out of a lung-CT for example and put thosefeatures in context by computing a description of possible malformations anddysfunctions.Our idea is good fitted for a hackathon as it is scalable in it’s size. The hyper-paramters, data size and the amount of diseases can be tested until the end ofthe hackathon.