Where we’re going, we won’t need real images — Q4 papers

After my first post on medical image segmentation papers (here’s the link), I worked on an update with mostly Q4 papers, based on Google Scholar research and arXiv CS surfing. I hope it’ll be useful to you!

1. A Survey on Deep Learning in Medical Image Analysis, Litjens et al. This has to be the most efficient way to start looking into this field to date. It’s a comprehensive review of more than 300 papers from the last 5 years and it clearly outlines how different models have come and gone. A perfect starter, also on the different use cases of deep learning in the medical field (i.e., object detection, organ segmentation, lesion detection, image registration, content-based image retrieval, etc.). “Groups and researchers that obtain good performance when applying deep learning algorithms often differentiate themselves in aspects outside of the deep net-
   work, like novel data preprocessing or augmentation techniques”, they write.
2. DeepRadiologyNet: Radiologist Level Pathology Detection in CT Head Images, Merkow et al. A much discussed paper I include not because of the actual goal (an automatic diagnosis tool for head CTs) since I’m more interested in specialised uses, but because of the size of their dataset, which is at 30.000 studies apparently. The results and their presentation are critically discussed in this post by John Zech. They are as well in this blogpost by Luke Oakden-Rayner (@DrLukeOR on twitter), both very much recommended.
3. Y-net: 3D intracranial artery segmentation using a convolutional autoencoder, Chen et al. An interesting contribution from the substructure department, so to speak. Using MRAs, a CAE does automated segmentation of intracranial arteries, which could be useful in other organs as well, the liver for example. It performs well compared to traditional methods.
4. Deep convolutional neural networks for segmenting 3D in vivo multiphoton images of vasculature in Alzheimer disease mouse models, Haft-Javaherian et al. This one is equally impressive, concerning vasculature detection. They announce to be above SOA as well as above human expert quality.
5. Detection-aided liver lesion segmentation using deep learning, Bellver et al. A neat paper from the Machine learning 4 Health workshop at NIPS, doing unsurprisingly lesion detection in liver tissue. The lesion detection is done after the segmentation step. Although from this year’s NIPS, I find it interesting that the basic architectures used are still DRIU and ResNet, both somewhat older models.
6. Simulating Patho-realistic Ultrasound Images using Deep Generative Networks with Adversarial Learning, Tom, Sheet. An increasing number of papers is about GANs simulating images (cf. https://arxiv.org/abs/1712.07695) in order to train unsupervised models. It’s well known that the availability of medical images in general is low, but that the availability of ultrasound images is even lower, necessitating data exchange relationships with institutions, a considerable hurdle for a lot of teams. This development also points to a possible general trend where models get more important than data.

Image taken from [6] showing their ultrasound imagery creation pipeline.

Finally, here’s a great post on proper books on deep learning and medical imaging coming out in 2018, check it out.

This is it for this second installment of my medical image segmentation review. Let me know if there are papers or posts I missed, I will gladly edit them in.