Oct 14, 2019

UII Publishes Seven AI Breakthrough Results in the World’s Top Medical Imaging Event—MICCAI 2019

Scientists from Shanghai United Imaging Intelligence (UII) published seven papers in the 22nd International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI 2019), Shenzhen, China, Oct 13 to Oct 17, 2019. The works cover breakthroughs from artificial intelligence methodologies for segmentation, classification, image fusion and registration, to computer-assisted diagnosis and assessment applications.

1 Mapping the Brain’s Functional Networks

Mild cognitive impairment (MCI) is a pre-clinical stage between normal aging and dementia, and early intervention could help slow down the progression from MCI to dementia. Neural imaging studies for early MCI can identify subtle imaging markers not visible in normal radiological reading and help better understand the mechanisms of MCI.

Using the new graph convolutional long short term memory networks, UII researchers proposed an early MCI diagnosis algorithm based on functional MRI sequences. The algorithm analyzes the dynamic functional connectivity patterns using graph convolutional networks and not only identifies early MCI from healthy subjects but also reports the anatomical regions and functional connections that contribute to such abnormality. This work also proved that by predicting gender and age in a multi-task setting, more accurate and robust classification can be achieved, and each subject receives a unique report of his or her own brain activations.

2 Knee MR Image Segmentation

A novel loss function, called Gradient Harmonized Dice Loss (GHDL), was proposed for convolutional neural network (CNN)–based volumetric image segmentation. The objective of GHDL is to address the quantity imbalance between voxel classes and to focus on the examples that are difficult to segment, in training. GHDL is not only applicable for two class segmentation but also further generalized to multi-class segmentation. In this paper, GHDL was employed in a 3D CNN for multiple-object segmentation of MR knee joint images and validated on the public dataset SKI10 and 637 scans from a local hospital. The results showed that GHDL outperformed the popular loss functions, such as Dice loss and Focal loss, and achieved better knee joint segmentation.

3 Pneumothorax Segmentation and Diagnosis

Pneumothorax is a critical abnormality that shall be treated with higher priority. Computerized triage can automatically identify and prioritize such acute cases to improve outcomes of healthcare. In this paper, a deep-learning-based framework to automatically segment the pneumothorax in chest X-rays was developed for such a triage system. Since a large number of pixel-level annotations is commonly needed but difficult to obtain for deep learning models, a weakly-supervised framework was proposed, which allows for partial training data to be annotated at the image-level. The attention masks derived from an image-level classification model are then used as the pixel-level masks for further training. Because the attention masks are rough and may have errors, a spatial label smoothing regularization technique is developed to deal with the uncertainty of the attention masks in the training of segmentation model. Experimental results showed that the proposed weakly-supervised segmentation algorithm relieves the need of well-annotated data and yields satisfactory performance on pneumothorax segmentation.

4 Cognitive Impairment: Prediction, Localization, and Identification 

Mild Cognitive Impairment (MCI) is the prodromal stage of Alzheimer’s disease (AD), with a high incident rate converting to AD. Hence, it is critical to identify MCI patients who will convert to AD patients for early and effective treatment. Recently, many machine learning or deep learning-based methods have been proposed to first localize the pathology-related brain regions and then extract respective features for MCI progression diagnosis. However, the intrinsic relationship between pathological region localization and respective feature extraction was usually neglected. To address this issue, in this paper, UII proposed a novel iterative attention focusing strategy for joint pathological region localization and identification of progressive MCI (pMCI) from stable MCI (sMCI). The pathological regions can be iteratively localized, and the respective diagnosis performance is in turn improved. The proposed algorithm outperformed the state-of-the-art methods on prediction accuracy, while additionally providing a focused attention map on specific pathological locations related to MCI progression. This allows for more insights and better understanding of the progression of MCI to AD.

5 Drawing the Midline of the Brain

Traumatic brain injury (TBI) could cause brain midline shift (MLS), and measuring of which serves as a quantitative indication of its severity. To date, comprehensive and automated quantification of MLS is still challenging for many healthcare providers (i.e., emergency physicians). To address such issues, UII proposed a novel and automated method for the robust midline delineation, which is based on the regression and multi-scale feature integration strategy. Traditional methods usually extract the symmetry related features to delineate midlines. However, they are not robust for largely deformed brains or less symmetric hemispheres. The method proposed by UII could deal with these largely deformed brains and provide quantitative parameters, which are used to assist the diagnosis of related diseases.

6 Image Registration and Fusion for Ablation Guidance

Thermal ablation elevates the temperature (55–65 Celsius) of a focal zone in the tumor and induces irreversible cell injury and eventually causes tumor apoptosis and coagulative necrosis. Therefore, accurate targeting of the tumor area is critical for ablating tumor tissues only and leaving the surrounding healthy tissues intact, which relies on the registration between pre-procedural image and intra-procedural image. Such a task is challenging as the large appearance gap and deformation between two images. For tackling these challenges, UII proposed a synthesis and inpainting-based MR-CT registration method to effectively and efficiently overlay the pre-procedural MR image onto intra-procedural CT image. The algorithm takes about 3 seconds for intra-procedural registration and can facilitate the operation of liver tumor thermal ablation.

7 Super Resolution of MR Images

In clinical practice, magnetic resonance (MR) imaging is often scanned with large slice thickness or slice distance due to scanning time limits. The acquired images are thus anisotropic, with much lower inter-slice resolution than the intra-slice resolution. For better coverage of the organs of interest, multiple anisotropic scans, each focuses on a certain direction, are usually acquired. In this work, UII proposed a 3D deep learning-based super-resolution (SR) framework to reconstruct the isotropic high-resolution MR images from multiple anisotropic scans. A spatially sparse fidelity loss was designed to keep the intensities of known slices the same before and after the reconstruction, and an adversarial regularization was adopted to make sure that the entire reconstructed image owns consistent appearance perceptually. For the first time, this approach achieved SR from multiple anisotropic images, instead of a single one, without using any supervision using isotropic high-resolution images. Such unique contribution makes the algorithm highly feasible and applicable to many real clinical scenarios, especially multi-plane MR super resolution.