Leveraging Multi-Task Learning to Cope With Poor and Missing Labels of Mammograms

In breast cancer screening, binary classification of mammograms is a common task aiming to determine whether a case is malignant or benign. A Computer-Aided Diagnosis (CADx) system based on a trainable classifier requires clean data and labels coming from a confirmed diagnosis. Unfortunately, such labels are not easy to obtain in clinical practice, since the histopathological reports of biopsy may not be available alongside mammograms, while normal cases may not have an explicit follow-up confirmation. Such ambiguities result either in reducing the number of samples eligible for training or in a label uncertainty that may decrease the performances. In this work, we maximize the number of samples for training relying on multi-task learning. We design a deep-neural-network-based classifier yielding multiple outputs in one forward pass. The predicted classes include binary malignancy, cancer probability estimation, breast density, and image laterality. Since few samples have all classes available and confirmed, we propose to introduce the uncertainty related to the classes as a per-sample weight during training. Such weighting prevents updating the network's parameters when training on uncertain or missing labels. We evaluate our approach on the public INBreast and private datasets, showing statistically significant improvements compared to baseline and independent state-of-the-art approaches. Moreover, we use mammograms from Susan G. Komen Tissue Bank for fine-tuning, further demonstrating the ability to improve the performances in our multi-task learning setup from raw clinical data. We achieved the binary classification performance of AUC = 80.46 on our private dataset and AUC = 85.23 on the INBreast dataset.

Expression of CRY2 Gene in the Brain Is Related to Human Navigation

Navigation is a complex cognitive process. CRY2 gene has been proposed to play an important role in navigation behaviors in various non-human animal species. Utilizing a recently developed neuroimaging-transcriptomics approach, the present study reported a tentative link between the CRY2 gene and human navigation. Specifically, we showed a significant pattern similarity between CRY2 gene expression in the human brain and navigation-related neural activation in functional magnetic resonance imaging. To further illuminate the functionality of CRY2 in human navigation, we examined the correlation between CRY2 expression and various cognitive processes underlying navigation, and found high correlation of CRY2 expression with neural activity of multiple cognitive domains, particularly object and shape perception and spatial memory. Further analyses on the relation between the neural activity of human navigation and the expression maps of genes of two CRY2-related pathways, i.e., the magnetoreceptive and circadian-related functions, found a trend of correlation for the CLOCK gene, a core circadian regulator gene, suggesting that CRY2 may modulate human navigation through its role in circadian rhythm. This observation was further confirmed by a behavioral study where individuals with better circadian regularity in daily life showed better sense of direction. Taken together, our study presents the first neural evidence that links CRY2 with human navigation, possibly through the modulation of circadian rhythm.

Review and Prospect: Artificial Intelligence in Advanced Medical Imaging

Artificial intelligence (AI) as an emerging technology is gaining momentum in medical imaging. Recently, deep learning-based AI techniques have been actively investigated in medical imaging, and its potential applications range from data acquisition and image reconstruction to image analysis and understanding. In this review, we focus on the use of deep learning in image reconstruction for advanced medical imaging modalities including magnetic resonance imaging (MRI), computed tomography (CT), and positron emission tomography (PET). Particularly, recent deep learning-based methods for image reconstruction will be emphasized, in accordance with their methodology designs and performances in handling volumetric imaging data. It is expected that this review can help relevant researchers understand how to adapt AI for medical imaging and which advantages can be achieved with the assistance of AI.

Vascularity and Dynamic Contrast-Enhanced Breast Magnetic Resonance Imaging

Angiogenesis is a key step in the initiation and progression of an invasive breast cancer. High microvessel density by morphological characterization predicts metastasis and poor survival in women with invasive breast cancers. However, morphologic characterization is subject to variability and only can evaluate a limited portion of an invasive breast cancer. Consequently, breast Magnetic Resonance Imaging (MRI) is currently being evaluated to assess vascularity. Recently, through the new field of radiomics, dynamic contrast enhanced (DCE)-MRI is being used to evaluate vascular density, vascular morphology, and detection of aggressive breast cancer biology. While DCE-MRI is a highly sensitive tool, there are specific features that limit computational evaluation of blood vessels. These include (1) DCE-MRI evaluates gadolinium contrast and does not directly evaluate biology, (2) the resolution of DCE-MRI is insufficient for imaging small blood vessels, and (3) DCE-MRI images are very difficult to co-register. Here we review computational approaches for detection and analysis of blood vessels in DCE-MRI images and present some of the strategies we have developed for co-registry of DCE-MRI images and early detection of vascularization.

Characterisation of Children's Head Motion for Magnetic Resonance Imaging With and Without General Anaesthesia

Head motion is one of the major reasons for artefacts in Magnetic Resonance Imaging (MRI), which is especially challenging for children who are often intimidated by the dimensions of the MR scanner. In order to optimise the MRI acquisition for children in the clinical setting, insights into children's motion patterns are essential. In this work, we analyse motion data from 61 paediatric patients. We compare structural MRI data of children imaged with and without general anaesthesia (GA), all scanned using the same hybrid PET/MR scanner. We analyse several metrics of motion based on the displacement relative to a reference, decompose the transformation matrix into translation and rotation, as well as investigate whether different regions in the brain are affected differently by the children's motion. Head motion for children without GA was significantly higher, with a median of the mean displacements of 2.19 ± 0.93 mm (median ± standard deviation) during 41.7±7.5 min scans; however, even anaesthetised children showed residual head motion (mean displacement of 1.12±0.35 mm). For both patient groups translation along the z-axis (along the scanner bore) was significantly larger in absolute terms (GA / no GA: 0.87±0.29/0.92 ± 0.49 mm) compared to the other directions. Considering directionality, both patient groups were moving in negative z-direction and thus, out of the scanner. The awake children additionally showed significantly more nodding rotation (0.33±0.20°). In future studies as well as in the clinical setting, these predominant types of motion need to be taken into consideration to limit artefacts and reduce re-scans due to poor image quality.

Active Learning Performance in Labeling Radiology Images Is 90% Effective

To train artificial intelligence (AI) systems on radiology images, an image labeling step is necessary. Labeling for radiology images usually involves a human radiologist manually drawing a (polygonal) shape onto the image and attaching a word to it. As datasets are typically large, this task is repetitive, time-consuming, error-prone, and expensive. The AI methodology of active learning (AL) can assist human labelers by continuously sorting the unlabeled images in order of information gain and thus getting the labeler always to label the most informative image next. We find that after about 10%, depending on the dataset, of the images in a realistic dataset are labeled, virtually all the information content has been learnt and the remaining images can be automatically labeled. These images can then be checked by the radiologist, which is far easier and faster to do. In this way, the entire dataset is labeled with much less human effort. We introduce AL in detail and expose the effectiveness using three real-life datasets. We contribute five distinct elements to the standard AL workflow creating an advanced methodology.

Induction of Robust Future Liver Remnant Hypertrophy Before Hepatectomy With a Modified Liver Venous Deprivation Technique Using a Trans-venous Access for Hepatic Vein Embolization

Purpose: Hepatic and/or portal vein embolization are performed before hepatectomy for patients with insufficient future liver remnant and usually achieved with a trans-hepatic approach. The aim of the present study is to describe a modified trans-venous liver venous deprivation technique (mLVD), avoiding the potential risks and limitations of a percutaneous approach to hepatic vein embolization, and to assess the safety, efficacy, and surgical outcome after mLVD.Materials and Methods: Retrospective single-center institutional review board-approved study. From March 2016 to June 2019, consecutive oncologic patients with combined portal and hepatic vein embolization were included. CT volumetric analysis was performed before and after mLVD to assess liver hypertrophy. Complications related to mLVD and surgical outcome were obtained from medical records.Results: Thirty patients (62.7 ± 14.5 years old, 20 men) with liver metastasis (60%) or primary liver cancer (40%) underwent mLVD. Twenty-one patients (70%) had hepatic vein anatomic variants. Technical success of mLVD was 100%. Four patients had complications (three minor and one major). FLR hypertrophy was 64.2% ± 51.3% (mean ± SD). Twenty-four patients (80%) underwent the planned hepatectomy and no surgery was canceled as a consequence of mLVD complications or insufficient hypertrophy. Fifty percent of patients (12/24) had no or mild complications after surgery (Clavien-Dindo 0–II), and 45.8% (11/24) had more serious complications (Clavien-Dindo III–IV). Thirty-day mortality was 4.2% (1/24).Conclusion: mLVD is an effective method to induce FLR hypertrophy. This technique is applicable in a wide range of oncologic situations and in patients with complex right liver vein anatomy.

Brain Tumor Segmentation From Multi-Modal MR Images via Ensembling UNets

Glioma is a type of severe brain tumor, and its accurate segmentation is useful in surgery planning and progression evaluation. Based on different biological properties, the glioma can be divided into three partially-overlapping regions of interest, including whole tumor (WT), tumor core (TC), and enhancing tumor (ET). Recently, UNet has identified its effectiveness in automatically segmenting brain tumor from multi-modal magnetic resonance (MR) images. In this work, instead of network architecture, we focus on making use of prior knowledge (brain parcellation), training and testing strategy (joint 3D+2D), ensemble and post-processing to improve the brain tumor segmentation performance. We explore the accuracy of three UNets with different inputs, and then ensemble the corresponding three outputs, followed by post-processing to achieve the final segmentation. Similar to most existing works, the first UNet uses 3D patches of multi-modal MR images as the input. The second UNet uses brain parcellation as an additional input. And the third UNet is inputted by 2D slices of multi-modal MR images, brain parcellation, and probability maps of WT, TC, and ET obtained from the second UNet. Then, we sequentially unify the WT segmentation from the third UNet and the fused TC and ET segmentation from the first and the second UNets as the complete tumor segmentation. Finally, we adopt a post-processing strategy by labeling small ET as non-enhancing tumor to correct some false-positive ET segmentation. On one publicly-available challenge validation dataset (BraTS2018), the proposed segmentation pipeline yielded average Dice scores of 91.03/86.44/80.58% and average 95% Hausdorff distances of 3.76/6.73/2.51 mm for WT/TC/ET, exhibiting superior segmentation performance over other state-of-the-art methods. We then evaluated the proposed method on the BraTS2020 training data through five-fold cross validation, with similar performance having also been observed. The proposed method was finally evaluated on 10 in-house data, the effectiveness of which has been established qualitatively by professional radiologists.

A Case-Based Review of Cerebral Venous Infarcts With Perfusion Imaging and Comparison to Arterial Ischemic Stroke

Cerebral venous thrombosis (CVT) and cerebral venous infarcts (CVI) are diagnostic dilemmas secondary to their rarity, non-specific symptomatology at presentation, and variable imaging features. Despite its relatively infrequence, CVT is particularly prevalent in the younger adult population and is a potentially life-threatening disease with devastating neurological complications if not addressed in a timely manner. However, when treated promptly, CVT has the potential for a more reversible course and favorable prognosis than arterial ischemic strokes (AIS). The pathophysiology of CVI is distinct from that of AIS and is closely related to its potentially reversible nature. Familiarity with the conventional and variant venous anatomy, as well as the temporal evolution of imaging findings, is crucial in establishing diagnostic confidence. The use of MR perfusion imaging (MRP) and arterial spin-labeling (ASL) can potentially aid in the diagnosis of CVT/CVI via characterization of cerebral blood flow. The presence and extent of a cerebral perfusion deficit on either CT or MRI may play a role in clinical outcomes for patients with CVT, although future larger studies must be performed. This review presents a case-based overview focusing on the classic imaging characteristics of CVT and CVI in conjunction with bolus MRP and ASL findings in the adult population.

Meningioma Segmentation in T1-Weighted MRI Leveraging Global Context and Attention Mechanisms

Purpose: Meningiomas are the most common type of primary brain tumor, accounting for ~30% of all brain tumors. A substantial number of these tumors are never surgically removed but rather monitored over time. Automatic and precise meningioma segmentation is, therefore, beneficial to enable reliable growth estimation and patient-specific treatment planning.Methods: In this study, we propose the inclusion of attention mechanisms on top of a U-Net architecture used as backbone: (i) Attention-gated U-Net (AGUNet) and (ii) Dual Attention U-Net (DAUNet), using a three-dimensional (3D) magnetic resonance imaging (MRI) volume as input. Attention has the potential to leverage the global context and identify features' relationships across the entire volume. To limit spatial resolution degradation and loss of detail inherent to encoder–decoder architectures, we studied the impact of multi-scale input and deep supervision components. The proposed architectures are trainable end-to-end and each concept can be seamlessly disabled for ablation studies.Results: The validation studies were performed using a five-fold cross-validation over 600 T1-weighted MRI volumes from St. Olavs Hospital, Trondheim University Hospital, Norway. Models were evaluated based on segmentation, detection, and speed performances, and results are reported patient-wise after averaging across all folds. For the best-performing architecture, an average Dice score of 81.6% was reached for an F1-score of 95.6%. With an almost perfect precision of 98%, meningiomas smaller than 3 ml were occasionally missed hence reaching an overall recall of 93%.Conclusion: Leveraging global context from a 3D MRI volume provided the best performances, even if the native volume resolution could not be processed directly due to current GPU memory limitations. Overall, near-perfect detection was achieved for meningiomas larger than 3 ml, which is relevant for clinical use. In the future, the use of multi-scale designs and refinement networks should be further investigated. A larger number of cases with meningiomas below 3 ml might also be needed to improve the performance for the smallest tumors.