Topics of Interest

The scope of this workshop includes, but is not strictly limited to, the topics as follows.

Un/semi/weakly-supervised learning

Supervised learning with medical images requires annotations by board-certified physicians. A cost-effective solution to this is constructing partially or weakly annotated training set. If a large amount of data is available, unsupervised pre-training (i.e., unsupervised representation learning) followed by supervised transfer learning can be considered.

Active learning

Annotating an input data that a model perceived as uncertain is more beneficial to the model improvement than a randomly chosen input data. Choosing the most informative data for annotation can significantly reduce the annotation cost in the medical image domain until achieving a satisfactory performance.

Label refinement

For some diseases with weak visual appearance such as small lung nodules in chest X-ray images, physicians often fail to make consensus on determining the presence or location of the disease. In this case, it is difficult to define the ground truth, which results in the problem of noisy labels.

Domain adaptation

Performance degradation caused by a gap between training and test domains is a typical problem, especially in medical image analysis. In almost all situations, the source of training data is limited to several medical centers, so it is practically impossible to cover a wide variety of races or imaging devices. To solve this problem, the model should be adapted to the new environment using a small amount of target data.

Out-of-distribution detection

In medicine, an incorrect prediction is much more dangerous than in other applications since the diagnosis directly affects the patient's clinical outcome. For input data that is outside of the training data distribution, the model is likely to make an inaccurate prediction. It is safe for a model to reject such input data.

Multi-modal learning

Physicians sometimes diagnose a disease by considering multiple modality data, e.g., radiography images and sonograms. If the multiple modalities contain complementary information about the target disease, the diagnostic performance of a recognition model can be improved.

Classification, detection, segmentation

Classification, detection, and segmentation of the target are the three major visual recognition tasks in the medical image domain.

Image enhancement

While the radiation dose is a key factor to acquire a high-quality radiography image, too much radiation raises concern to patients' health condition. For this reason, image enhancement techniques such as noise reduction are important to enhance the low-quality images from low radiation dose.

Voxel representations

One major feature of medical images that is distinguished from natural images is 3D volumetric nature of some image modalities, such as CT and MR, are used. It is necessary to have architecture and learning method suitable for the 3D volume to efficiently learn the voxel representations.

Longitudinal study

Following up visual changes of a patient's lesion over time is very important for accurate diagnose of the disease and planning the treatment. However, it is not only difficult to accurately quantify the changes but also labor-intensive for physicians to follow up the patient's status manually.

Image to text

Physicians read medical images and record the findings in the form of a report. Automatically generating templates of radiology reports by a visual recognition model can make the physician's workflow more efficient.