Medical Imaging Meets NIPS: A summary
This year I attended and presented a poster at the Medical Imaging Meets NIPs workshop. The workshop focused on bringing together professionals from both the medical imaging and machine learning communities. Altogether there were eleven talks and two poster sessions. Here I’m going to recap some of the highlights of the workshop. Presentations and posters generally discussed segmentation, classification, and/or image reconstruction.
Segmentation
Before coming to this workshop I must admit that I did not fully understand the value of image segmentation. My thought process was always something along the lines: why would you just want to outline something in an image and not also classify it? This workshop changed my view on the value of segmentation.
Radiation Therapy
Raj Jena, a radiologist at Cambridge University and a Microsoft researcher, gave a presentation on “ Pixel Perfectionism — Machine learning and Adaptive Radiation Therapy.” In the talk he described how machine learning could help provide better treatments and optimize workflows. In order for patients to receive proper radiation therapy it is important to pinpoint the exact boundary of the tumor. By locating the exact boundary between the tumor and healthy tissue, treatments can deliver more radiation as there is less risk of damaging healthy tissue. However, currently, segmentation is done manually by radiologists. This often causes discrepancies between different radiologists, which can noticeably affect treatment results. Consistency is also important in gauging the effectiveness of drugs used in combination with radiation because if the radiation is not the same across the patients it is nearly impossible to tell if improvements are caused by the drug or better radiation.
Machine learning offers the opportunity to provide consistency and more accurate segmentation. Secondly, machine learning models can often run in seconds whereas radiologists often take several hours to manually segment images. This time can be better spent plotting the course of treatment or seeing additional patients. Jena also described how machine learning could allow him to become a “super radiation oncologist.”
ML can enable oncologists to both better adapt treatments to changes in the shape and size of healthy tissues and to help oncologists predict possible adverse affects of radiation therapy. For instance, Jena described how he is using simple methods such as Gaussian processes to predict potential side effects of radiation.
This was one of my favorite talks of the entire workshop and I urge you to check out Jena’s full presentation.
Building quality datasets
A common theme throughout the workshop was the quality of annotations and the difficulty building good medical imaging datasets. This is particularly true in segmentation tasks where a model can only be as good as its annotators and the annotators must be skilled radiologists themselves.
One possible way to accelerate the annotation process is through active learning. Tanveer Syeda-Mahmood of IBM briefly brought this up when discussing IBM’s work on radiology. With active learning one might start with a small labeled dataset and several expert human annotators. The ML algorithm learns the training set well enough so that it can annotate easy images itself and the experts annotate the hard edge cases. Specifically, images that the classifier scores below a threshold of certainty are then sent to humans to manually annotate. One of the posters (by Girro et al) also discussed using active learning to help effectively train a semantic image segmentation network.
Active learning may address part of the problem, however it does not entirely solve the quality issue. The central question is how can researchers develop a accurate dataset when even the experts disagree on boundaries. With respect to this point, Bjorne Menze presented on the construction of the BraTS dataset. The BraTS dataset is one of the largest brain imaging datasets. He fused data from several different annotators in order to create the “ground truth.” Since its creation BraTS has held several different challenges. One of the challenges involved segmenting all the tumors with machine learning algorithms and the most recent focused on predicting overall survival.
Localization, detection, and classification
Accurately classifying diseases found in medical images was a prominent topic at the workshop. Detecting objects/ROIs and accurately classifying them is a challenging task in medical imaging. This is largely due to the variety of modalities (and dimensions) of medical images (i.e. X-Ray MRI, CT, Ultrasound, and Sonogram), the size of the images, and (as with segmentation) limited annotated (and sometimes low quality) training data. As such, presenters showcased an interesting variety of techniques for overcoming these obstacles.
Ivana Igsum discussed deep learning techniques in cardiac imaging. In particular, she described her work in accurately detecting calcification in arteries. She described how her and her team developed methods to automatically score calcium and categorize cardiovascular disease risk. To do this her team used a multi-layer CNN approach.
Later in the day Yaroslav Nikulin presented on the winning approach from the digital mammography challenge.
Posters
Natalia Antropova, Benjamin Huynh and Maryellen Giger of the University of Chicago had an interesting poster on using an LSTM to perform breast DCI-MRI classification. This involved inputting 3d MRI images from multiple time steps after a contrast dye was applied. They then extracted features from these images using a CNN which they fed to the LSTM and outputted a prediction. Altogether this poster provided an interesting application of a LSTM (and CNN) to handle “4d” medical imaging data.
My poster focused on my current work in progress with respect to using object detectors to accurately localize and classify multiple conditions in Chest X-Rays. My major goals are to investigate how well object detection algorithms perform on a limited dataset as opposed to multi-label classification CNNs trained on the entire dataset. I think object detectors have a lot of potential at localizing and classifying diseases/conditions in medical images if configured and trained properly, however they are limited by the shortage of labeled bounding box data, which is one of the reasons I found the following poster very interesting.
Hiba Chougrad and Hamid Zouaki had an interesting poster on transfer learning for breast imaging classification. In the abstract Convolutional Neural Networks for Breast Cancer Screening: Transfer Learning with Exponential Decay, they described testing several different transfer learning methods. For example, they compared fine tuning and utilizing a CNN pretrained on imagenet with randomly initializing weights. In the end they discovered the optimal technique was to use an exponentially decaying learning rate to fine tune the layers. So for the bottom layers (i.e., the ones closest to the softmax), the learning rate would be the highest and for the upper layers the learning rate would be the lowest. This intuitively makes a lot sense as the bottom layers tend to learn the most dataset relevant features. By using these and related techniques we can (hopefully) develop accurate models without having large datasets.
Reconstruction and generation
Heartflow
I usually am not impressed with industrial pitches touting how great their product is and how it will “revolutionize [insert industry].” However, Heartflow and their DeepLumen blood vessel segmentation algorithm definitely impressed me. The product reduced unnecessary angiograms by 83% and is FDA approved. I will not go into extended detail here, but I think Heartflow is a good example of machine learning having an impact in a real world environment.
Two of the other presenters also touched on reconstruction as well. Igsum’s talk (previously mentioned) discussed a method for constructing a routine CT from a low dose CT. Daniel Rueckert of Imperial College described how ML based reconstruction could enable more time stamps in imaging.
Posters
One of the poster that I found particularly interesting was MR-to-CT Synthesis using Cycle-Consistent Generative Adversarial Networks. In this work the authors (Wolterink et al.) take the popular CycleGAN algorithm and used it to convert MRI images into CT images. This is a potentially very useful application that could prevent patients from having to have multiple imaging procedures. Additionally, CTs expose patients to radiation so this could also potentially reduce radiation exposure.
There were also a posters by Virdi et al. on Synthetic Medical Images from Dual Generative Adversarial Networks and Mardani et al. on Deep Generative Adversarial Networks for Compressed Sensing (GANCS) Automates MRI.
Tools and platforms
Several speakers spoke about new tools aimed at making medical image analysis with machine learning more accessible to both clinicians and ML researchers. Jorge Cardoso described NiftyNet and how it enables researchers to develop medical imaging models more easily. NiftyNet is built on Tensorflow and it includes many simple to use modules for loading high dimensional inputs.
Also on the tools side, G. Varoquax presented on NILearn, a Python module for Neuro-Imaging data, built on top of scikit-learn. Just as scikit-learn seeks to make ML accessible to people with basic programming skills, the goal of NILearn is to do the same with brain imaging. The only systems related poster was from Makkie and Liu from the University of Georgia. It focused on their brain initiative platform for neuro-imaging and how it fused several different technologies including Spark, AWS and Tensorflow. Finally, the DLTK toolkit had their own poster at the conference. Altogether there were some really interesting toolkits and platforms that should help make medical image analysis with machine learning more accessible to everyone.
Other talks
Wiro Nessen had an interesting presentation on the confluence of biomedical imaging and genetic data. In the talk he described how both large genetic and imaging datasets could be combined to try to detect bio-markers in the images. The synthesis of the two areas could also help detect diseases early and figure out more targeted treatments.
Announcements
- At the end of her talk, Ivana Igsum announced that the first Deep Learning in Medical Imaging or MIDL event is taking place in Amsterdam in July.
- I’m continuously adding new papers, conferences, and tools to my machine learning healthcare curated list. However, it’s a big job so make a PR and contribute today!
- I’m starting a new Slack channel dedicated to machine learning in healthcare. So if you are interested feel free to join.
- My summary of Machine Learning for Healthcare or ML4H at NIPs will be out in the next few weeks. This will cover everything from hospital operations (like LOS forecasting and hand hygiene), to mining electronic medical records, to drug discovery, to analyzing genomic data. So stay tuned.
Comments and questions are welcome as always.
