On the challenges and perspectives of foundation models for medical image analysis

Medical Image Analysis, Год журнала: 2023, Номер 91, С. 102996 - 102996

Опубликована: Окт. 12, 2023

Язык: Английский

---

Current and Emerging Trends in Medical Image Segmentation With Deep Learning

IEEE Transactions on Radiation and Plasma Medical Sciences, Год журнала: 2023, Номер 7(6), С. 545 - 569

Опубликована: Апрель 10, 2023

In recent years, the segmentation of anatomical or pathological structures using deep learning has experienced a widespread interest in medical image analysis. Remarkably successful performance been reported many imaging modalities and for variety clinical contexts to support clinicians computer-assisted diagnosis, therapy, surgical planning purposes. However, despite increasing amount challenges, there remains little consensus on which methodology performs best. Therefore, we examine this article numerous developments breakthroughs brought since rise U-Net-inspired architectures. Especially, focus technical challenges emerging trends that community is now focusing on, including conditional generative adversarial cascaded networks, Transformers, contrastive learning, knowledge distillation, active prior embedding, cross-modality multistructure analysis, federated semi-supervised self-supervised paradigms. We also suggest possible avenues be further investigated future research efforts.

Язык: Английский

---

DRAC 2022: A public benchmark for diabetic retinopathy analysis on ultra-wide optical coherence tomography angiography images

Patterns, Год журнала: 2024, Номер 5(3), С. 100929 - 100929

Опубликована: Фев. 8, 2024

We described a challenge named "DRAC - Diabetic Retinopathy Analysis Challenge" in conjunction with the 25th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI 2022). Within this challenge, we provided DRAC datset, an ultra-wide optical coherence tomography angiography (UW-OCTA) dataset (1,103 images), addressing three primary clinical tasks: diabetic retinopathy (DR) lesion segmentation, image quality assessment, DR grading. The scientific community responded positively to 11, 12, 13 teams submitting different solutions for these tasks, respectively. This paper presents concise summary analysis of top-performing results across all tasks. These could provide practical guidance developing accurate classification segmentation models assessment diagnosis using UW-OCTA images, potentially improving diagnostic capabilities healthcare professionals. has been released support development computer-aided systems evaluation.

Язык: Английский

---

Hybrid Masked Image Modeling for 3D Medical Image Segmentation

IEEE Journal of Biomedical and Health Informatics, Год журнала: 2024, Номер 28(4), С. 2115 - 2125

Опубликована: Янв. 30, 2024

Masked image modeling (MIM) with transformer backbones has recently been exploited as a powerful self-supervised pre-training technique. The existing MIM methods adopt the strategy to mask random patches of and reconstruct missing pixels, which only considers semantic information at lower level, causes long time. This paper presents HybridMIM, novel hybrid learning method based on masked for 3D medical segmentation. Specifically, we design two-level masking hierarchy specify how in sub-volumes are masked, effectively providing constraints higher level information. Then learn images three levels, including: 1) partial region prediction key contents image, largely reduces time burden (pixel-level); 2) patch-masking perception spatial relationship between each sub-volume (region-level); 3) drop-out-based contrastive samples within mini-batch, further improves generalization ability framework (sample-level). proposed is versatile support both CNN encoder backbones, also enables pre-train decoders We conduct comprehensive experiments five widely-used public segmentation datasets, including BraTS2020, BTCV, MSD Liver, Spleen, BraTS2023. experimental results show clear superiority HybridMIM against competing supervised methods, approaches, other terms quantitative metrics, speed performance qualitative observations. codes available https://github.com/ge-xing/HybridMIM .

Язык: Английский

---

AI-driven 3D bioprinting for regenerative medicine: From bench to bedside

Bioactive Materials, Год журнала: 2024, Номер 45, С. 201 - 230

Опубликована: Ноя. 23, 2024

Язык: Английский

---

Multi-site, Multi-domain Airway Tree Modeling

Medical Image Analysis, Год журнала: 2023, Номер 90, С. 102957 - 102957

Опубликована: Сен. 9, 2023

Язык: Английский

---

SwinUNETR-V2: Stronger Swin Transformers with Stagewise Convolutions for 3D Medical Image Segmentation

Lecture notes in computer science, Год журнала: 2023, Номер unknown, С. 416 - 426

Опубликована: Янв. 1, 2023

Язык: Английский

---

COSST: Multi-Organ Segmentation With Partially Labeled Datasets Using Comprehensive Supervisions and Self-Training

IEEE Transactions on Medical Imaging, Год журнала: 2024, Номер 43(5), С. 1995 - 2009

Опубликована: Янв. 15, 2024

Deep learning models have demonstrated remarkable success in multi-organ segmentation but typically require large-scale datasets with all organs of interest annotated. However, medical image are often low sample size and only partially labeled, i.e., a subset Therefore, it is crucial to investigate how learn unified model on the available labeled leverage their synergistic potential. In this paper, we systematically partial-label problem theoretical empirical analyses prior techniques. We revisit from perspective partial label supervision signals identify two derived ground truth one pseudo labels. propose novel two-stage framework termed COSST, which effectively efficiently integrates comprehensive self-training. Concretely, first train an initial using truth-based then iteratively incorporate signal To mitigate performance degradation caused by unreliable labels, assess reliability labels via outlier detection latent space exclude most each self-training iteration. Extensive experiments conducted public three private tasks over 12 CT datasets. Experimental results show that our proposed COSST achieves significant improvement baseline method, individual networks trained dataset. Compared state-of-the-art methods, demonstrates consistent superior various different training data sizes.

Язык: Английский

---

MedLSAM: Localize and segment anything model for 3D CT images

Medical Image Analysis, Год журнала: 2024, Номер 99, С. 103370 - 103370

Опубликована: Окт. 15, 2024

Язык: Английский

---

One-Shot Weakly-Supervised Segmentation in 3D Medical Images

IEEE Transactions on Medical Imaging, Год журнала: 2023, Номер 43(1), С. 175 - 189

Опубликована: Июль 13, 2023

Deep neural networks typically require accurate and a large number of annotations to achieve outstanding performance in medical image segmentation. One-shot weakly-supervised learning are promising research directions that reduce labeling effort by new class from only one annotated using coarse labels instead, respectively. In this work, we present an innovative framework for 3D segmentation with one-shot settings. Firstly propagation-reconstruction network is proposed propagate scribbles volume unlabeled images based on the assumption anatomical patterns different human bodies similar. Then multi-level similarity denoising module designed refine embeddings anatomical- pixel-level. After expanding pseudo masks, observe miss-classified voxels mainly occur at border region propose extract self-support prototypes specific refinement. Based these results, further train model noisy label training strategy. Experiments three CT MRI datasets show method obtains significant improvement over state-of-the-art methods performs robustly even under severe imbalance low contrast. Code publicly available https://github.com/LWHYC/OneShot_WeaklySeg.

Язык: Английский

---