How can we test AI performance? This question seems trivial, but it isn't. Standard benchmarks often have problems such as in-distribution and small-size test sets, oversimplified metrics, unfair comparisons, and short-term outcome pressure. As a consequence, good performance on standard benchmarks does not guarantee success in real-world scenarios. To address these problems, we present Touchstone, a large-scale collaborative segmentation benchmark of 9 types of abdominal organs. This benchmark is based on 5,195 training CT scans from 76 hospitals around the world and 5,903 testing CT scans from 11 additional hospitals. This diverse test set enhances the statistical significance of benchmark results and rigorously evaluates AI algorithms across various out-of-distribution scenarios. We invited 14 inventors of 19 AI algorithms to train their algorithms, while our team, as a third party, independently evaluated these algorithms on three test sets. In addition, we also evaluated pre-existing AI frameworks--which, differing from algorithms, are more flexible and can support different algorithms--including MONAI from NVIDIA, nnU-Net from DKFZ, and numerous other open-source frameworks. We are committed to expanding this benchmark to encourage more innovation of AI algorithms for the medical domain.

Interactive segmentation, an integration of AI algorithms and human expertise, premises to improve the accuracy and efficiency of curating large-scale, detailed-annotated datasets in healthcare. Human experts revise the annotations predicted by AI, and in turn, AI improves its predictions by learning from these revised annotations. This interactive process continues to enhance the quality of annotations until no major revision is needed from experts. The key challenge is how to leverage AI predicted and expert revised annotations to iteratively improve the AI. Two problems arise: (1) The risk of catastrophic forgetting--the AI tends to forget the previously learned classes if it is only retrained using the expert revised classes. (2) Computational inefficiency when retraining the AI using both AI predicted and expert revised annotations; moreover, given the dominant AI predicted annotations in the dataset, the contribution of newly revised annotations--often account for a very small fraction--to the AI training remains marginal. This paper proposes Continual Tuning to address the problems from two perspectives: network design and data reuse. Firstly, we design a shared network for all classes followed by class-specific networks dedicated to individual classes. To mitigate forgetting, we freeze the shared network for previously learned classes and only update the class-specific network for revised classes. Secondly, we reuse a small fraction of data with previous annotations to avoid over-computing. The selection of such data relies on the importance estimate of each data. The importance score is computed by combining the uncertainty and consistency of AI predictions. Our experiments demonstrate that Continual Tuning achieves a speed 16x greater than repeatedly training AI from scratch without compromising the performance.

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of Gemini models in cross-modal reasoning and language understanding will enable a wide variety of use cases and we discuss our approach toward deploying them responsibly to users.

Prior to the deep learning era, shape was commonly used to describe the objects. Nowadays, state-of-the-art (SOTA) algorithms in medical imaging are predominantly diverging from computer vision, where voxel grids, meshes, point clouds, and implicit surface models are used. This is seen from numerous shape-related publications in premier vision conferences as well as the growing popularity of ShapeNet (about 51,300 models) and Princeton ModelNet (127,915 models). For the medical domain, we present a large collection of anatomical shapes (e.g., bones, organs, vessels) and 3D models of surgical instrument, called MedShapeNet, created to facilitate the translation of data-driven vision algorithms to medical applications and to adapt SOTA vision algorithms to medical problems. As a unique feature, we directly model the majority of shapes on the imaging data of real patients. As of today, MedShapeNet includes 23 dataset with more than 100,000 shapes that are paired with annotations (ground truth). Our data is freely accessible via a web interface and a Python application programming interface (API) and can be used for discriminative, reconstructive, and variational benchmarks as well as various applications in virtual, augmented, or mixed reality, and 3D printing. Exemplary, we present use cases in the fields of classification of brain tumors, facial and skull reconstructions, multi-class anatomy completion, education, and 3D printing. In future, we will extend the data and improve the interfaces. The project pages are: https://medshapenet.ikim.nrw/ and https://github.com/Jianningli/medshapenet-feedback

This study leverages synthetic data as a validation set to reduce overfitting and ease the selection of the best model in AI development. While synthetic data have been used for augmenting the training set, we find that synthetic data can also significantly diversify the validation set, offering marked advantages in domains like healthcare, where data are typically limited, sensitive, and from out-domain sources (i.e., hospitals). In this study, we illustrate the effectiveness of synthetic data for early cancer detection in computed tomography (CT) volumes, where synthetic tumors are generated and superimposed onto healthy organs, thereby creating an extensive dataset for rigorous validation. Using synthetic data as validation can improve AI robustness in both in-domain and out-domain test sets. Furthermore, we establish a new continual learning framework that continuously trains AI models on a stream of out-domain data with synthetic tumors. The AI model trained and validated in dynamically expanding synthetic data can consistently outperform models trained and validated exclusively on real-world data. Specifically, the DSC score for liver tumor segmentation improves from 26.7% (95% CI: 22.6%-30.9%) to 34.5% (30.8%-38.2%) when evaluated on an in-domain dataset and from 31.1% (26.0%-36.2%) to 35.4% (32.1%-38.7%) on an out-domain dataset. Importantly, the performance gain is particularly significant in identifying very tiny liver tumors (radius < 5mm) in CT volumes, with Sensitivity improving from 33.1% to 55.4% on an in-domain dataset and 33.9% to 52.3% on an out-domain dataset, justifying the efficacy in early detection of cancer. The application of synthetic data, from both training and validation perspectives, underlines a promising avenue to enhance AI robustness when dealing with data from varying domains.

Creating large-scale and well-annotated datasets to train AI algorithms is crucial for automated tumor detection and localization. However, with limited resources, it is challenging to determine the best type of annotations when annotating massive amounts of unlabeled data. To address this issue, we focus on polyps in colonoscopy videos and pancreatic tumors in abdominal CT scans; both applications require significant effort and time for pixel-wise annotation due to the high dimensional nature of the data, involving either temporary or spatial dimensions. In this paper, we develop a new annotation strategy, termed Drag&Drop, which simplifies the annotation process to drag and drop. This annotation strategy is more efficient, particularly for temporal and volumetric imaging, than other types of weak annotations, such as per-pixel, bounding boxes, scribbles, ellipses, and points. Furthermore, to exploit our Drag&Drop annotations, we develop a novel weakly supervised learning method based on the watershed algorithm. Experimental results show that our method achieves better detection and localization performance than alternative weak annotations and, more importantly, achieves similar performance to that trained on detailed per-pixel annotations. Interestingly, we find that, with limited resources, allocating weak annotations from a diverse patient population can foster models more robust to unseen images than allocating per-pixel annotations for a small set of images. In summary, this research proposes an efficient annotation strategy for tumor detection and localization that is less accurate than per-pixel annotations but useful for creating large-scale datasets for screening tumors in various medical modalities.

Annotating medical images, particularly for organ segmentation, is laborious and time-consuming. For example, annotating an abdominal organ requires an estimated rate of 30-60 minutes per CT volume based on the expertise of an annotator and the size, visibility, and complexity of the organ. Therefore, publicly available datasets for multi-organ segmentation are often limited in data size and organ diversity. This paper proposes an active learning method to expedite the annotation process for organ segmentation and creates the largest multi-organ dataset (by far) with the spleen, liver, kidneys, stomach, gallbladder, pancreas, aorta, and IVC annotated in 8,448 CT volumes, equating to 3.2 million slices. The conventional annotation methods would take an experienced annotator up to 1,600 weeks (or roughly 30.8 years) to complete this task. In contrast, our annotation method has accomplished this task in three weeks (based on an 8-hour workday, five days a week) while maintaining a similar or even better annotation quality. This achievement is attributed to three unique properties of our method: (1) label bias reduction using multiple pre-trained segmentation models, (2) effective error detection in the model predictions, and (3) attention guidance for annotators to make corrections on the most salient errors. Furthermore, we summarize the taxonomy of common errors made by AI algorithms and annotators. This allows for continuous revision of both AI and annotations and significantly reduces the annotation costs required to create large-scale datasets for a wider variety of medical imaging tasks.

Image synthesis approaches, e.g., generative adversarial networks, have been popular as a form of data augmentation in medical image analysis tasks. It is primarily beneficial to overcome the shortage of publicly accessible data and associated quality annotations. However, the current techniques often lack control over the detailed contents in generated images, e.g., the type of disease patterns, the location of lesions, and attributes of the diagnosis. In this work, we adapt the latest advance in the generative model, i.e., the diffusion model, with the added control flow using lesion-specific visual and textual prompts for generating dermatoscopic images. We further demonstrate the advantage of our diffusion model-based framework over the classical generation models in both the image quality and boosting the segmentation performance on skin lesions. It can achieve a 9% increase in the SSIM image quality measure and an over 5% increase in Dice coefficients over the prior arts.

An increasing number of public datasets have shown a marked impact on automated organ segmentation and tumor detection. However, due to the small size and partially labeled problem of each dataset, as well as a limited investigation of diverse types of tumors, the resulting models are often limited to segmenting specific organs/tumors and ignore the semantics of anatomical structures, nor can they be extended to novel domains. To address these issues, we propose the CLIP-Driven Universal Model, which incorporates text embedding learned from Contrastive Language-Image Pre-training (CLIP) to segmentation models. This CLIP-based label encoding captures anatomical relationships, enabling the model to learn a structured feature embedding and segment 25 organs and 6 types of tumors. The proposed model is developed from an assembly of 14 datasets, using a total of 3,410 CT scans for training and then evaluated on 6,162 external CT scans from 3 additional datasets. We rank first on the Medical Segmentation Decathlon (MSD) public leaderboard and achieve state-of-the-art results on Beyond The Cranial Vault (BTCV). Additionally, the Universal Model is computationally more efficient (6x faster) compared with dataset-specific models, generalized better to CT scans from varying sites, and shows stronger transfer learning performance on novel tasks.

Early detection and localization of pancreatic cancer can increase the 5-year survival rate for patients from 8.5% to 20%. Artificial intelligence (AI) can potentially assist radiologists in detecting pancreatic tumors at an early stage. Training AI models require a vast number of annotated examples, but the availability of CT scans obtaining early-stage tumors is constrained. This is because early-stage tumors may not cause any symptoms, which can delay detection, and the tumors are relatively small and may be almost invisible to human eyes on CT scans. To address this issue, we develop a tumor synthesis method that can synthesize enormous examples of small pancreatic tumors in the healthy pancreas without the need for manual annotation. Our experiments demonstrate that the overall detection rate of pancreatic tumors, measured by Sensitivity and Specificity, achieved by AI trained on synthetic tumors is comparable to that of real tumors. More importantly, our method shows a much higher detection rate for small tumors. We further investigate the per-voxel segmentation performance of pancreatic tumors if AI is trained on a combination of CT scans with synthetic tumors and CT scans with annotated large tumors at an advanced stage. Finally, we show that synthetic tumors improve AI generalizability in tumor detection and localization when processing CT scans from different hospitals. Overall, our proposed tumor synthesis method has immense potential to improve the early detection of pancreatic cancer, leading to better patient outcomes.