In data-scarce scenarios, deep learning models often overfit to noise and irrelevant patterns, which limits their ability to generalize to unseen samples. To address these challenges in medical image segmentation, we introduce Diff-UMamba, a novel architecture that combines the UNet framework with the mamba mechanism for modeling long-range dependencies. At the heart of Diff-UMamba is a Noise Reduction Module (NRM), which employs a signal differencing strategy to suppress noisy or irrelevant activations within the encoder. This encourages the model to filter out spurious features and enhance task-relevant representations, thereby improving its focus on clinically meaningful regions. As a result, the architecture achieves improved segmentation accuracy and robustness, particularly in low-data settings. Diff-UMamba is evaluated on multiple public datasets, including MSD (lung and pancreas) and AIIB23, demonstrating consistent performance gains of 1-3% over baseline methods across diverse segmentation tasks. To further assess performance under limited-data conditions, additional experiments are conducted on the BraTS-21 dataset by varying the proportion of available training samples. The approach is also validated on a small internal non-small cell lung cancer (NSCLC) dataset for gross tumor volume (GTV) segmentation in cone beam CT (CBCT), where it achieves a 4-5% improvement over the baseline.

The Diffusion Probabilistic Model (DPM) has demonstrated remarkable performance across a variety of generative tasks. The inherent randomness in diffusion models helps address issues such as blurring at the edges of medical images and labels, positioning Diffusion Probabilistic Models (DPMs) as a promising approach for lesion segmentation. However, we find that the current training and inference strategies of diffusion models result in an uneven distribution of attention across different timesteps, leading to longer training times and suboptimal solutions. To this end, we propose UniSegDiff, a novel diffusion model framework designed to address lesion segmentation in a unified manner across multiple modalities and organs. This framework introduces a staged training and inference approach, dynamically adjusting the prediction targets at different stages, forcing the model to maintain high attention across all timesteps, and achieves unified lesion segmentation through pre-training the feature extraction network for segmentation. We evaluate performance on six different organs across various imaging modalities. Comprehensive experimental results demonstrate that UniSegDiff significantly outperforms previous state-of-the-art (SOTA) approaches. The code is available at https://github.com/HUYILONG-Z/UniSegDiff.

Leveraging the powerful capabilities of diffusion models has yielded quite effective results in medical image segmentation tasks. However, existing methods typically transfer the original training process directly without specific adjustments for segmentation tasks. Furthermore, the commonly used pre-trained diffusion models still have deficiencies in feature extraction. Based on these considerations, we propose LEAF, a medical image segmentation model grounded in latent diffusion models. During the fine-tuning process, we replace the original noise prediction pattern with a direct prediction of the segmentation map, thereby reducing the variance of segmentation results. We also employ a feature distillation method to align the hidden states of the convolutional layers with the features from a transformer-based vision encoder. Experimental results demonstrate that our method enhances the performance of the original diffusion model across multiple segmentation datasets for different disease types. Notably, our approach does not alter the model architecture, nor does it increase the number of parameters or computation during the inference phase, making it highly efficient.

To compare the predictive value of minimal ablative margin (MAM) quantification using tumor segmentation on intraprocedural contrast-enhanced hepatic arterial (HAP) versus portal venous phase (PVP) CT on local outcomes following percutaneous thermal ablation of colorectal liver metastases (CRLM). This dual-center retrospective study included patients undergoing thermal ablation of CRLM with intraprocedural preablation and postablation contrast-enhanced CT imaging between 2009 and 2021. Tumors were segmented in both HAP and PVP CT phases using an artificial intelligence-based auto-segmentation model and reviewed by a trained radiologist. The MAM was quantified using a biomechanical deformable image registration process. The area under the receiver operating characteristic curve (AUROC) was used to compare the prognostic value for predicting local tumor progression (LTP). Among 81 patients (60 y±13, 53 men), 151 CRLMs were included. During 29.4 months of median follow-up, LTP was noted in 24/151 (15.9%). Median tumor volumes on HAP and PVP CT were 1.7 mL and 1.2 mL, respectively, with respective median MAMs of 2.3 and 4.0 mm (both P< 0.001). The AUROC for 1-year LTP prediction was 0.78 (95% CI: 0.70-0.85) on HAP and 0.84 (95% CI: 0.78-0.91) on PVP (P= 0.002). During CT-guided percutaneous thermal ablation, MAM measured based on tumors segmented on PVP images conferred a higher predictive accuracy of ablation outcomes among CRLM patients than those segmented on HAP images, supporting the use of PVP rather than HAP images for segmentation during ablation of CRLMs.

Single-photon emission computed tomography (SPECT) plays a crucial role in detecting bone metastases from lung cancer. However, its low spatial resolution and lesion similarity to benign structures present significant challenges for accurate segmentation, especially for lesions of varying sizes. We propose a deep learning-based segmentation framework that integrates conditional adversarial learning with a multi-scale feature extraction generator. The generator employs cascade dilated convolutions, multi-scale modules, and deep supervision, while the discriminator utilizes multi-scale L1 loss computed on image-mask pairs to guide segmentation learning. The proposed model was evaluated on a dataset of 286 clinically annotated SPECT scintigrams. It achieved a Dice Similarity Coefficient (DSC) of 0.6671, precision of 0.7228, and recall of 0.6196 - outperforming both classical and recent adversarial segmentation models in multi-scale lesion detection, especially for small and clustered lesions. Our results demonstrate that the integration of multi-scale feature learning with adversarial supervision significantly improves the segmentation of bone metastasis in SPECT imaging. This approach shows potential for clinical decision support in the management of lung cancer.

Many deep learning methods have been proposed for brain tumor segmentation from multi-modal Magnetic Resonance Imaging (MRI) scans that are important for accurate diagnosis and treatment planning. However, supervised learning needs a large amount of labeled data to perform well, where the time-consuming and expensive annotation process or small size of training set will limit the model's performance. To deal with these problems, self-supervised pre-training is an appealing solution due to its feature learning ability from a set of unlabeled images that is transferable to downstream datasets with a small size. However, existing methods often overlook the utilization of multi-modal information and multi-scale features. Therefore, we propose a novel Self-Supervised Learning (SSL) framework that fully leverages multi-modal MRI scans to extract modality-invariant features for brain tumor segmentation. First, we employ a Siamese Block-wise Modality Masking (SiaBloMM) strategy that creates more diverse model inputs for image restoration to simultaneously learn contextual and modality-invariant features. Meanwhile, we proposed Overlapping Random Modality Sampling (ORMS) to sample voxel pairs with multi-scale features for self-distillation, enhancing voxel-wise representation which is important for segmentation tasks. Experiments on the BraTS 2024 adult glioma segmentation dataset showed that with a small amount of labeled data for fine-tuning, our method improved the average Dice by 3.80 percentage points. In addition, when transferred to three other small downstream datasets with brain tumors from different patient groups, our method also improved the dice by 3.47 percentage points on average, and outperformed several existing SSL methods. The code is availiable at https://github.com/HiLab-git/Vox-MMSD.

Pancreatic cancer carries a poor prognosis and relies on endoscopic ultrasound (EUS) for targeted biopsy and radiotherapy. However, the speckle noise, low contrast, and unintuitive appearance of EUS make segmentation of pancreatic tumors with fully supervised deep learning (DL) models both error-prone and dependent on large, expert-curated annotation datasets. To address these challenges, we present TextSAM-EUS, a novel, lightweight, text-driven adaptation of the Segment Anything Model (SAM) that requires no manual geometric prompts at inference. Our approach leverages text prompt learning (context optimization) through the BiomedCLIP text encoder in conjunction with a LoRA-based adaptation of SAM's architecture to enable automatic pancreatic tumor segmentation in EUS, tuning only 0.86% of the total parameters. On the public Endoscopic Ultrasound Database of the Pancreas, TextSAM-EUS with automatic prompts attains 82.69% Dice and 85.28% normalized surface distance (NSD), and with manual geometric prompts reaches 83.10% Dice and 85.70% NSD, outperforming both existing state-of-the-art (SOTA) supervised DL models and foundation models (e.g., SAM and its variants). As the first attempt to incorporate prompt learning in SAM-based medical image segmentation, TextSAM-EUS offers a practical option for efficient and robust automatic EUS segmentation. Code is available at https://github.com/HealthX-Lab/TextSAM-EUS .

In data-scarce scenarios, deep learning models often overfit to noise and irrelevant patterns, which limits their ability to generalize to unseen samples. To address these challenges in medical image segmentation, we introduce Diff-UMamba, a novel architecture that combines the UNet framework with the mamba mechanism to model long-range dependencies. At the heart of Diff-UMamba is a noise reduction module, which employs a signal differencing strategy to suppress noisy or irrelevant activations within the encoder. This encourages the model to filter out spurious features and enhance task-relevant representations, thereby improving its focus on clinically significant regions. As a result, the architecture achieves improved segmentation accuracy and robustness, particularly in low-data settings. Diff-UMamba is evaluated on multiple public datasets, including medical segmentation decathalon dataset (lung and pancreas) and AIIB23, demonstrating consistent performance gains of 1-3% over baseline methods in various segmentation tasks. To further assess performance under limited data conditions, additional experiments are conducted on the BraTS-21 dataset by varying the proportion of available training samples. The approach is also validated on a small internal non-small cell lung cancer dataset for the segmentation of gross tumor volume in cone beam CT, where it achieves a 4-5% improvement over baseline.

Pancreatic cancer carries a poor prognosis and relies on endoscopic ultrasound (EUS) for targeted biopsy and radiotherapy. However, the speckle noise, low contrast, and unintuitive appearance of EUS make segmentation of pancreatic tumors with fully supervised deep learning (DL) models both error-prone and dependent on large, expert-curated annotation datasets. To address these challenges, we present TextSAM-EUS, a novel, lightweight, text-driven adaptation of the Segment Anything Model (SAM) that requires no manual geometric prompts at inference. Our approach leverages text prompt learning (context optimization) through the BiomedCLIP text encoder in conjunction with a LoRA-based adaptation of SAM's architecture to enable automatic pancreatic tumor segmentation in EUS, tuning only 0.86% of the total parameters. On the public Endoscopic Ultrasound Database of the Pancreas, TextSAM-EUS with automatic prompts attains 82.69% Dice and 85.28% normalized surface distance (NSD), and with manual geometric prompts reaches 83.10% Dice and 85.70% NSD, outperforming both existing state-of-the-art (SOTA) supervised DL models and foundation models (e.g., SAM and its variants). As the first attempt to incorporate prompt learning in SAM-based medical image segmentation, TextSAM-EUS offers a practical option for efficient and robust automatic EUS segmentation.

Recent advances in deep learning have led to robust automated tools for segmentation of abdominal computed tomography (CT). Meanwhile, segmentation of magnetic resonance imaging (MRI) is substantially more challenging due to the inherent signal variability and the increased effort required for annotating training datasets. Hence, existing approaches are trained on limited sets of MRI sequences, which might limit their generalizability. To characterize the landscape of MRI abdominal segmentation tools, we present here a comprehensive benchmarking of the three state-of-the-art and open-source models: MRSegmentator, MRISegmentator-Abdomen, and TotalSegmentator MRI. Since these models are trained using labor-intensive manual annotation cycles, we also introduce and evaluate ABDSynth, a SynthSeg-based model purely trained on widely available CT segmentations (no real images). More generally, we assess accuracy and generalizability by leveraging three public datasets (not seen by any of the evaluated methods during their training), which span all major manufacturers, five MRI sequences, as well as a variety of subject conditions, voxel resolutions, and fields-of-view. Our results reveal that MRSegmentator achieves the best performance and is most generalizable. In contrast, ABDSynth yields slightly less accurate results, but its relaxed requirements in training data make it an alternative when the annotation budget is limited. The evaluation code and datasets are given for future benchmarking at https://github.com/deepakri201/AbdoBench, along with inference code and weights for ABDSynth.