Subcortical segmentation of the fetal brain in 3D ultrasound using deep learning.

The quantification of subcortical volume development from 3D fetal ultrasound can provide important diagnostic information during pregnancy monitoring. However, manual segmentation of subcortical structures in ultrasound volumes is time-consuming and challenging due to low soft tissue contrast, speckle and shadowing artifacts. For this reason, we developed a convolutional neural network (CNN) for the automated segmentation of the choroid plexus (CP), lateral posterior ventricle horns (LPVH), cavum septum pellucidum et vergae (CSPV), and cerebellum (CB) from 3D ultrasound. As ground-truth labels are scarce and expensive to obtain, we applied few-shot learning, in which only a small number of manual annotations (n = 9) are used to train a CNN. We compared training a CNN with only a few individually annotated volumes versus many weakly labelled volumes obtained from atlas-based segmentations. This showed that segmentation performance close to intra-observer variability can be obtained with only a handful of manual annotations. Finally, the trained models were applied to a large number (n = 278) of ultrasound image volumes of a diverse, healthy population, obtaining novel US-specific growth curves of the respective structures during the second trimester of gestation.

PMC-LLaMA: toward building open-source language models for medicine.

OBJECTIVE: Recently, large language models (LLMs) have showcased remarkable capabilities in natural language understanding. While demonstrating proficiency in everyday conversations and question-answering (QA) situations, these models frequently struggle in domains that require precision, such as medical applications, due to their lack of domain-specific knowledge. In this article, we describe the procedure for building a powerful, open-source language model specifically designed for medicine applications, termed as PMC-LLaMA. MATERIALS AND METHODS: We adapt a general-purpose LLM toward the medical domain, involving data-centric knowledge injection through the integration of 4.8M biomedical academic papers and 30K medical textbooks, as well as comprehensive domain-specific instruction fine-tuning, encompassing medical QA, rationale for reasoning, and conversational dialogues with 202M tokens. RESULTS: While evaluating various public medical QA benchmarks and manual rating, our lightweight PMC-LLaMA, which consists of only 13B parameters, exhibits superior performance, even surpassing ChatGPT. All models, codes, and datasets for instruction tuning will be released to the research community. DISCUSSION: Our contributions are 3-fold: (1) we build up an open-source LLM toward the medical domain. We believe the proposed PMC-LLaMA model can promote further development of foundation models in medicine, serving as a medical trainable basic generative language backbone; (2) we conduct thorough ablation studies to demonstrate the effectiveness of each proposed component, demonstrating how different training data and model scales affect medical LLMs; (3) we contribute a large-scale, comprehensive dataset for instruction tuning. CONCLUSION: In this article, we systematically investigate the process of building up an open-source medical-specific LLM, PMC-LLaMA.

Sensorless volumetric reconstruction of fetal brain freehand ultrasound scans with deep implicit representation.

Three-dimensional (3D) ultrasound imaging has contributed to our understanding of fetal developmental processes by providing rich contextual information of the inherently 3D anatomies. However, its use is limited in clinical settings, due to the high purchasing costs and limited diagnostic practicality. Freehand 2D ultrasound imaging, in contrast, is routinely used in standard obstetric exams, but inherently lacks a 3D representation of the anatomies, which limits its potential for more advanced assessment. Such full representations are challenging to recover even with external tracking devices due to internal fetal movement which is independent from the operator-led trajectory of the probe. Capitalizing on the flexibility offered by freehand 2D ultrasound acquisition, we propose ImplicitVol to reconstruct 3D volumes from non-sensor-tracked 2D ultrasound sweeps. Conventionally, reconstructions are performed on a discrete voxel grid. We, however, employ a deep neural network to represent, for the first time, the reconstructed volume as an implicit function. Specifically, ImplicitVol takes a set of 2D images as input, predicts their locations in 3D space, jointly refines the inferred locations, and learns a full volumetric reconstruction. When testing natively-acquired and volume-sampled 2D ultrasound video sequences collected from different manufacturers, the 3D volumes reconstructed by ImplicitVol show significantly better visual and semantic quality than the existing interpolation-based reconstruction approaches. The inherent continuity of implicit representation also enables ImplicitVol to reconstruct the volume to arbitrarily high resolutions. As formulated, ImplicitVol has the potential to integrate seamlessly into the clinical workflow, while providing richer information for diagnosis and evaluation of the developing brain.

Knowledge-enhanced visual-language pre-training on chest radiology images.

While multi-modal foundation models pre-trained on large-scale data have been successful in natural language understanding and vision recognition, their use in medical domains is still limited due to the fine-grained nature of medical tasks and the high demand for domain knowledge. To address this challenge, we propose an approach called Knowledge-enhanced Auto Diagnosis (KAD) which leverages existing medical domain knowledge to guide vision-language pre-training using paired chest X-rays and radiology reports. We evaluate KAD on four external X-ray datasets and demonstrate that its zero-shot performance is not only comparable to that of fully supervised models but also superior to the average of three expert radiologists for three (out of five) pathologies with statistical significance. Moreover, when few-shot annotation is available, KAD outperforms all existing approaches in fine-tuning settings, demonstrating its potential for application in different clinical scenarios.

Self-Supervised Tumor Segmentation With Sim2Real Adaptation.

This paper targets on self-supervised tumor segmentation. We make the following contributions: (i) we take inspiration from the observation that tumors are often characterised independently of their contexts, we propose a novel proxy task "layer-decomposition", that closely matches the goal of the downstream task, and design a scalable pipeline for generating synthetic tumor data for pre-training; (ii) we propose a two-stage Sim2Real training regime for unsupervised tumor segmentation, where we first pre-train a model with simulated tumors, and then adopt a self-training strategy for downstream data adaptation; (iii) when evaluating on different tumor segmentation benchmarks, e.g. BraTS2018 for brain tumor segmentation and LiTS2017 for liver tumor segmentation, our approach achieves state-of-the-art segmentation performance under the unsupervised setting. While transferring the model for tumor segmentation under a low-annotation regime, the proposed approach also outperforms all existing self-supervised approaches; (iv) we conduct extensive ablation studies to analyse the critical components in data simulation, and validate the necessity of different proxy tasks. We demonstrate that, with sufficient texture randomization in simulation, model trained on synthetic data can effortlessly generalise to datasets with real tumors.

Learning to map 2D ultrasound images into 3D space with minimal human annotation.

In fetal neurosonography, aligning two-dimensional (2D) ultrasound scans to their corresponding plane in the three-dimensional (3D) space remains a challenging task. In this paper, we propose a convolutional neural network that predicts the position of 2D ultrasound fetal brain scans in 3D atlas space. Instead of purely supervised learning that requires heavy annotations for each 2D scan, we train the model by sampling 2D slices from 3D fetal brain volumes, and target the model to predict the inverse of the sampling process, resembling the idea of self-supervised learning. We propose a model that takes a set of images as input, and learns to compare them in pairs. The pairwise comparison is weighted by the attention module based on its contribution to the prediction, which is learnt implicitly during training. The feature representation for each image is thus computed by incorporating the relative position information to all the other images in the set, and is later used for the final prediction. We benchmark our model on 2D slices sampled from 3D fetal brain volumes at 18-22 weeks' gestational age. Using three evaluation metrics, namely, Euclidean distance, plane angles and normalized cross correlation, which account for both the geometric and appearance discrepancy between the ground-truth and prediction, in all these metrics, our model outperforms a baseline model by as much as 23%, when the number of input images increases. We further demonstrate that our model generalizes to (i) real 2D standard transthalamic plane images, achieving comparable performance as human annotations, as well as (ii) video sequences of 2D freehand fetal brain scans.

Low-Memory CNNs Enabling Real-Time Ultrasound Segmentation Towards Mobile Deployment.

Convolutional Neural Networks (CNNs), which are currently state-of-the-art for most image analysis tasks, are ill suited to leveraging the key benefits of ultrasound imaging - specifically, ultrasound's portability and real-time capabilities. CNNs have large memory footprints, which obstructs their implementation on mobile devices, and require numerous floating point operations, which results in slow CPU inference times. In this paper, we propose three approaches to training efficient CNNs that can operate in real-time on a CPU (catering to the clinical setting), with a low memory footprint, for minimal compromise in accuracy. We first demonstrate the power of 'thin' CNNs, with very few feature channels, for fast medical image segmentation. We then leverage separable convolutions to further speed up inference, reduce parameter count and facilitate mobile deployment. Lastly, we propose a novel knowledge distillation technique to boost the accuracy of light-weight models, while maintaining inference speed-up. For a negligible sacrifice in test set Dice performance on the challenging ultrasound analysis task of nerve segmentation, our final proposed model processes images at 30fps on a CPU, which is 9× faster than the standard U-Net, while requiring 420× less space in memory.

VP-Nets : Efficient automatic localization of key brain structures in 3D fetal neurosonography.

Three-dimensional (3D) fetal neurosonography is used clinically to detect cerebral abnormalities and to assess growth in the developing brain. However, manual identification of key brain structures in 3D ultrasound images requires expertise to perform and even then is tedious. Inspired by how sonographers view and interact with volumes during real-time clinical scanning, we propose an efficient automatic method to simultaneously localize multiple brain structures in 3D fetal neurosonography. The proposed View-based Projection Networks (VP-Nets), uses three view-based Convolutional Neural Networks (CNNs), to simplify 3D localizations by directly predicting 2D projections of the key structures onto three anatomical views. While designed for efficient use of data and GPU memory, the proposed VP-Nets allows for full-resolution 3D prediction. We investigated parameters that influence the performance of VP-Nets, e.g. depth and number of feature channels. Moreover, we demonstrate that the model can pinpoint the structure in 3D space by visualizing the trained VP-Nets, despite only 2D supervision being provided for a single stream during training. For comparison, we implemented two other baseline solutions based on Random Forest and 3D U-Nets. In the reported experiments, VP-Nets consistently outperformed other methods on localization. To test the importance of loss function, two identical models are trained with binary corss-entropy and dice coefficient loss respectively. Our best VP-Net model achieved prediction center deviation: 1.8 ±â€¯1.4 mm, size difference: 1.9 ±â€¯1.5 mm, and 3D Intersection Over Union (IOU): 63.2 ±â€¯14.7% when compared to the ground truth. To make the whole pipeline intervention free, we also implement a skull-stripping tool using 3D CNN, which achieves high segmentation accuracy. As a result, the proposed processing pipeline takes a raw ultrasound brain image as input, and output a skull-stripped image with five detected key brain structures.

Fully-automated alignment of 3D fetal brain ultrasound to a canonical reference space using multi-task learning.

Methods for aligning 3D fetal neurosonography images must be robust to (i) intensity variations, (ii) anatomical and age-specific differences within the fetal population, and (iii) the variations in fetal position. To this end, we propose a multi-task fully convolutional neural network (FCN) architecture to address the problem of 3D fetal brain localization, structural segmentation, and alignment to a referential coordinate system. Instead of treating these tasks as independent problems, we optimize the network by simultaneously learning features shared within the input data pertaining to the correlated tasks, and later branching out into task-specific output streams. Brain alignment is achieved by defining a parametric coordinate system based on skull boundaries, location of the eye sockets, and head pose, as predicted from intracranial structures. This information is used to estimate an affine transformation to align a volumetric image to the skull-based coordinate system. Co-alignment of 140 fetal ultrasound volumes (age range: 26.0 ±â€¯4.4 weeks) was achieved with high brain overlap and low eye localization error, regardless of gestational age or head size. The automatically co-aligned volumes show good structural correspondence between fetal anatomies.

Ω-Net (Omega-Net): Fully automatic, multi-view cardiac MR detection, orientation, and segmentation with deep neural networks.

Pixelwise segmentation of the left ventricular (LV) myocardium and the four cardiac chambers in 2-D steady state free precession (SSFP) cine sequences is an essential preprocessing step for a wide range of analyses. Variability in contrast, appearance, orientation, and placement of the heart between patients, clinical views, scanners, and protocols makes fully automatic semantic segmentation a notoriously difficult problem. Here, we present Ω-Net (Omega-Net): A novel convolutional neural network (CNN) architecture for simultaneous localization, transformation into a canonical orientation, and semantic segmentation. First, an initial segmentation is performed on the input image; second, the features learned during this initial segmentation are used to predict the parameters needed to transform the input image into a canonical orientation; and third, a final segmentation is performed on the transformed image. In this work, Ω-Nets of varying depths were trained to detect five foreground classes in any of three clinical views (short axis, SA; four-chamber, 4C; two-chamber, 2C), without prior knowledge of the view being segmented. This constitutes a substantially more challenging problem compared with prior work. The architecture was trained using three-fold cross-validation on a cohort of patients with hypertrophic cardiomyopathy (HCM, N=42) and healthy control subjects (N=21). Network performance, as measured by weighted foreground intersection-over-union (IoU), was substantially improved for the best-performing Ω-Net compared with U-Net segmentation without localization or orientation (0.858 vs 0.834). In addition, to be comparable with other works, Ω-Net was retrained from scratch using five-fold cross-validation on the publicly available 2017 MICCAI Automated Cardiac Diagnosis Challenge (ACDC) dataset. The Ω-Net outperformed the state-of-the-art method in segmentation of the LV and RV bloodpools, and performed slightly worse in segmentation of the LV myocardium. We conclude that this architecture represents a substantive advancement over prior approaches, with implications for biomedical image segmentation more generally.